go-v.orgdoi

Augmented Intelligence for Fault Diagnosis in Advanced Manufacturing

Cohen, Joseph · Deep Blue (University of Michigan) · 2023

  • DOI 10.7302/7331
  • Publisher University of Michigan
  • Type dissertation
  • Issued 2023

Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.

Deep Blue (University of Michigan) · University of Michigan · dissertation · 2023

Abstract

Artificial intelligence and machine learning methods have demonstrated immense potential in the advanced manufacturing research community with the advent of the Fourth Industrial Revolution by substantially reducing life-cycle costs coupled with reduced downtime. To facilitate advancements in prognostics and health management, data-driven fault diagnosis has emerged as a key area of focus. However, several challenges remain when implementing established artificial intelligence methods for fault diagnosis.

For example, conventional supervised machine learning approaches rely on large, labeled training sets that are often difficult to obtain in practice, and involve black-box models that lack explainability. Based on fundamental guiding principles such as accessibility, computational efficiency, explainability, reliability, and robustness, industrial augmented intelligence techniques are proposed to address the gaps in the state of the art for advanced manufacturing. The techniques developed vary based on the level of supervision and explainability, with an emphasis of semi-supervised learning present throughout.

The first approach handles the problem of diagnosing synchronization faults in timed event systems using programmable logic controller (PLC) signal data. Conventionally used to control and time increasingly complex asynchronous system operations, there has been relatively little research on directly using these signals as a reliable data source for detecting and diagnosing event synchronization faults. An industrial augmented intelligence (iAI) method that utilizes Timed Petri Nets to select inherently explainable time delay features is developed for this fault diagnosis problem.

This hybrid method combines the utility of discrete event models with machine learning approaches, with successful applications in unsupervised, semi-supervised, and fully supervised settings. Key highlights include the development of three unique and novel heuristics utilized to generate a Timed Petri Net of a nominal operating condition given input PLC signals, with successful validation performed on a use case study in semiconductor manufacturing. The second application focuses on semi-supervised clustering of weakly labeled, highly imbalanced, and high-dimensional in-process measurement data.

A new semi-supervised anomaly classification scheme is proposed based on partially labeled datasets, signifying explainability derived from targeted human assistance. The approach involves finding intersections between samples that are considered outliers to generally unlabeled data as well as inliers to known fault data, with additional functionality to recommend potential new faults and borderline cases. This approach has demonstrated excellent capability in detecting local as well as global anomalies, with a 65% reduction in false positives compared to conventional semi-supervised anomaly detection approaches.

Lastly, the thesis develops an explainable AI method employing Shapley-based clustering analysis for fault diagnosis and prognosis extended for semi-supervised learning. Validated on two case studies, including an important benchmark on engine prognostics, the method focuses on deriving explanations from developed black-box model predictions.

The prognostics methodology, expanded to include predictions for current health state and forecasting eventual failing component(s), results in a 38% reduction in remaining useful life root-mean-square error (RMSE) compared to previously published results of the same dataset split and area under precision recall (AUPR) curve scores of forecasted component-level failures exceeding 0.95. Furthermore, Shapley values are explored as a transformational tool to derive human-explainable clusters that are later characterized via simple and insightful decision rules comprised of 1-2 variables in terms of their original scale.

The developed industrial augmented intelligence methods empower users by synergizing data and human expertise as available to provide actionable solutions to some of the relevant fault diagnosis challenges present in modern advanced manufacturing.

Open the resolved locator · 10.7302/7331

Metadata licence: OpenAlex metadata (CC0)