The Library
Building transformative framework for isolation and mitigation of quality defects in multi-station assembly systems using deep learning
Tools
Sinha, Sumit (2021) Building transformative framework for isolation and mitigation of quality defects in multi-station assembly systems using deep learning. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Sinha_2021.pdf - Submitted Version - Requires a PDF viewer. Download (43Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3755529
Abstract
The manufacturing industry is undergoing significant transformation towards electrification (e-mobility). This transformation has intensified critical development of new lightweight materials, structures and assembly processes supporting high volume and high variety production of Battery Electric Vehicles (BEVs). As new materials and processes get developed it is crucial to address quality defects detection, prediction, and prevention especially given that e-mobility products interlink quality and safety, for example, assembly of ‘live’ battery systems. These requirements necessitate the development of methodologies that ensure quality requirements of products are satisfied from Job 1. This means ensuring high right-first-time ratio during process design by reducing manual and ineffective trial-and-error process adjustments; and, then continuing this by maintaining near zero-defect manufacturing during production by reducing Mean-Time-to-Detection and Mean-Time-to-Resolution for critical quality defects. Current technologies for isolating and mitigating quality issues provide limited performance within complex manufacturing systems due to (i) limited modelling abilities and lack capabilities to leverage point cloud quality monitoring data provided by recent measurement technologies such as 3D scanners to isolate defects; (ii) extensive dependence on manual expertise to mitigate the isolated defects; and, (iii) lack of integration between data-driven and physics-based models resulting in limited industrial applicability, scalability and interpretability capabilities, hence constitute a significant barrier towards ensuring quality requirements throughout the product lifecycle.
The study develops a transformative framework that goes beyond improving the accuracy and performance of current approaches and overcomes fundamental barriers for isolation and mitigation of product shape error quality defects in multi-station assembly systems (MAS). The proposed framework is based on three methodologies which explore MAS: (i) response to quality defects by isolating process parameters (root causes (RCs)) causing unaccepted shape error defects; (ii) correction of the isolated RCs by determining corrective actions (CA) policy to mitigate unaccepted shape error defects; and, (iii) training, scalability and interpretability of (i) and (ii) by establishing closed-loop in-process (CLIP) capability that integrates in-line point cloud data, deep learning approaches of (i) and (ii) and physics-based models to provide comprehensive data-driven defect identification and RC isolation (causality analysis). The developed methodologies include:
(i) Object Shape Error Response (OSER) to isolate RCs within single- and multi-station assembly systems (OSER-MAS) by developing Bayesian 3D-convolutional neural network architectures that process point cloud data and are trained using physics-based models and have capabilities to relate complex product shape error patterns to RCs. It quantifies uncertainties and is applicable during the design phase when no quality monitoring data is available.
(ii) Object Shape Error Correction (OSEC) to generate CAs that mitigate RCs and simultaneously account for cost and quality key performance indicators (KPIs), MAS reconfigurability, and stochasticity by developing a deep reinforcement learning framework that estimates effective and feasible CAs without manual expertise.
(iii) Closed-Loop In-Process (CLIP) to enable industrial adoption of approaches (i) & (ii) by firstly enhancing the scalability by using (a) closed-loop training, and (b) continual/transfer learning. This is important as training deep learning models for a MAS is time-intensive and requires large amounts of labelled data; secondly providing interpretability and transparency for the estimated RCs that drive costly CAs using (c) 3D gradient-based class activation maps.
The methods are implemented as independent kernels and then integrated within a transformative framework which is further verified, validated, and benchmarked using industrial-scale automotive sheet metal assembly case studies such as car door and cross-member. They demonstrate 29% better performance for RC isolation and 40% greater effectiveness for CAs than current statistical and engineering-based approaches.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > Q Science (General) Q Science > QA Mathematics > QA76 Electronic computers. Computer science. Computer software T Technology > TL Motor vehicles. Aeronautics. Astronautics T Technology > TS Manufactures |
||||
Library of Congress Subject Headings (LCSH): | Electric vehicles, Neural networks (Computer science), Machine learning, Deep learning (Machine learning), Reinforcement learning, Assembly-line methods, Production engineering, Engineering inspection, Automobiles -- Defects, Manufactures -- Defects | ||||
Official Date: | December 2021 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Warwick Manufacturing Group | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Ceglarek, Darek ; Franciosa, Pasquale | ||||
Sponsors: | Warwick Manufacturing Group ; Engineering and Physical Sciences Research Council ; UK Research and Innovation (Agency) | ||||
Format of File: | |||||
Extent: | xvii, 296 leaves : illustrations | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |
Downloads
Downloads per month over past year