Predicting shim gaps in aircraft assembly with machine learning and sparse sensing

Krithika Manohar, Thomas A. Hogan, Jim Buttrick, Ashis G. Banerjee, J. Nathan Kutz, Steven L. BruntonJournal of Manufacturing Systems 2018 · DOI

TL;DR

This paper introduces PIXI-DUST, a novel machine learning approach that drastically reduces the number of measurements needed to predict shim gaps in aircraft assembly. By leveraging robust principal component analysis (RPCA) and sparse sensor optimization, the method achieves high prediction accuracy using only ~3% of typical laser scan data, promising significant time and cost savings in aircraft production.

What Problem Does This Paper Address?

Modern aircraft assembly involves fitting thousands of components, and manufacturing tolerances often lead to small gaps between them. These gaps must be filled with custom-made "shims" to ensure the aircraft's structural integrity and aerodynamic performance. Traditionally, this process is labor-intensive and time-consuming, involving manual gap measurement and shimming, or computationally intensive analysis of vast amounts of metrology data. This bottleneck significantly delays production. The paper aims to address this by developing a data-driven method that can accurately predict shim gaps using a minimal set of measurements, thereby accelerating the assembly process.

Methods

The core of the PIXI-DUST approach lies in two key components:

  • Robust Principal Component Analysis (RPCA): To extract meaningful patterns from historical shim gap data, the authors employ RPCA. This technique decomposes the high-dimensional measurement data into a low-rank component (representing dominant, coherent structures across aircraft) and a sparse component (representing outliers and noise). This robust approach is crucial for handling real-world, potentially noisy, manufacturing data. The low-rank component's principal components (features) capture the underlying, shared patterns in shim distributions.
  • Sparse Sensor Optimization: Once the dominant features are identified, the paper uses a greedy algorithm based on pivoted QR factorization to select an optimized subset of measurement locations (sensors). This selection aims to maximize the information gained from each measurement, effectively identifying the most informative points for reconstructing the entire shim gap profile. This is framed as a D-optimal experimental design problem, minimizing the volume of the confidence ellipsoid for the reconstructed gaps.

The overall workflow involves an "offline training" phase where RPCA and sensor optimization are performed on historical data, and an "online execution" phase where only the optimized measurements are collected and used for prediction on new aircraft.

Datasets

The study utilizes historical production data from 54 representative Boeing commercial aircraft. This dataset consists of 10076 laser gap measurements for each aircraft. The data is used for both training the RPCA model and optimizing the sensor locations, and for rigorous cross-validation. The authors do not introduce a new benchmark dataset but rather leverage existing, real-world production data.

Key Insights and Ideas

  • Aircraft Shim Gaps Exhibit Low-Rank Structure: The fundamental assumption that shim gap distributions across different aircraft share common, low-dimensional patterns is validated by the RPCA analysis. This suggests that the complex problem of shimming can be simplified by understanding these underlying features.
  • Optimized Sparse Sensing is Highly Effective: The paper demonstrates that a small, strategically chosen set of measurement points can capture nearly as much information as a full scan, drastically reducing data acquisition time and computational load.
  • RPCA is Crucial for Robustness: The use of RPCA, rather than standard PCA, is vital for handling the inevitable outliers and noise present in manufacturing data, leading to more reliable feature extraction.
  • Segmentation Improves Accuracy: Breaking down the aircraft into smaller, individual shim segments for analysis further improves prediction accuracy and reduces the number of required sensors for each segment.

Key Results

  • High Prediction Accuracy: The PIXI-DUST algorithm successfully predicts 99% of shim gaps within the desired measurement tolerance of ±0.005 inches when considering the entire aircraft as a single structure.
  • Drastic Data Reduction: The method achieves this accuracy using only around 3% of the total laser scan points (specifically, 100 optimized sensors out of 10076 total measurements).
  • Segmented Improvement: When analyzing individual shim segments, prediction accuracy reaches 96-99% for most segments, with optimal sensor usage ranging from 2% to 6% of the points within that segment.
  • Robustness to Noise: The RPCA component effectively handles outliers, ensuring that the extracted features are not unduly influenced by erroneous measurements.

Limitations and Future Work

  • Generalizability to Other Metrology: While the method is demonstrated on laser scan data, its applicability to other metrology equipment (like point clouds) is mentioned, but not fully explored. For point clouds, the reduction in acquisition time might be less significant, but computational load reduction remains a key benefit.
  • Outlier Information: The authors note that the sparse outliers removed by RPCA might contain valuable information about anomalies and defects, suggesting future work in using machine learning to classify these outliers.
  • Manufacturing Process Variations: The study observed that aircraft tend to cluster in the PCA feature space, hinting at potential differences in manufacturing processes. Mining this data and using PCA coefficients for process control is proposed as a future direction.
  • Online Learning: The paper mentions ongoing work to implement the technology for online learning on future aircraft, which would allow the model to adapt and improve as more data becomes available.
  • Transfer Learning: Investigating how knowledge acquired for one shim location can be transferred to similar parts is also suggested.

Summary updated

Abstract

A modern aircraft may require on the order of thousands of custom shims to fill gaps between structural components in the airframe that arise due to manufacturing tolerances adding up across large structures. These shims, whether liquid or solid, are necessary to eliminate gaps, maintain structural performance, and minimize pull-down forces required to bring the aircraft into engineering nominal configuration for peak aerodynamic efficiency. Currently, gap filling is a time-consuming process, involving either expensive by-hand inspection or computations on vast quantities of measurement data from increasingly sophisticated metrology equipment. In either case, this amounts to significant delays in production, with much of the time being spent in the critical path of the aircraft assembly. In this work, we present an alternative strategy for predictive shimming, based on machine learning and sparse sensing to first learn gap distributions from historical data, and then design optimized sparse sensing strategies to streamline the collection and processing of data. This new approach is based on the assumption that patterns exist in shim distributions across aircraft, and that these patterns may be mined and used to reduce the burden of data collection and processing in future aircraft. Specifically, robust principal component analysis is used to extract low-dimensional patterns in the gap measurements while rejecting outliers. Next, optimized sparse sensors are obtained that are most informative about the dimensions of a new aircraft in these low-dimensional principal components. We demonstrate the success of the proposed approach, known within Boeing as PIXel Identification Despite Uncertainty in Sensor Technology (PIXI-DUST), on historical production data from 54 representative Boeing commercial aircraft. Our algorithm successfully predicts 99% of the shim gaps within the desired measurement tolerance using around 3% of the laser scan points that are typically required; all results are rigorously cross-validated.
BibTeX
@article{manohar2018predicting,
  author = {Krithika Manohar and Thomas A. Hogan and Jim Buttrick and Ashis G. Banerjee and J. Nathan Kutz and Steven L. Brunton},
  title = {{Predicting shim gaps in aircraft assembly with machine learning and sparse sensing}},
  journal = {Journal of Manufacturing Systems},
  year = {2018},
  doi = {10.1016/j.jmsy.2018.01.011},
}
Figure