Improving manufacturing operations using deep learning: Multi-faceted research in monitoring, diagnosis, and optimization

Today's manufacturing is rapidly marching towards a new era of quality, efficiency, and sustainability, driven by advances in sensing, computing, and machine learning. While data analytics, particularly machine learning, has been increasingly applied in manufacturing, new methodologies are still critically needed to fully harness manufacturing big data. This dissertation presents a suite of novel machine learning, particularly deep learning, methods to address challenges in various manufacturing processes and scenarios, encompassing diagnosis, monitoring, planning, and optimization. The contributions of this dissertation are briefly summarized as follows. A deep learning-based fault diagnosis method was developed for potentially mixed faults in multiple parts of rotating machinery. Deep learning models are employed to dissect the mixed failure modes into isolated components, which encompass various severity levels, locations, and operating conditions. This work addressed a large combination of 48 fault types, surpassing the capabilities of existing methods available at the time of publication. The research also explored the robustness of the classifiers to Gaussian noise and the detection of new, unlearned fault types. A Kalman filter-based approach for high-speed video multi-object tracking with limited and occasionally missing information, extending existing intersection-over-union methods. This approach outperformed leading methods available at the time of publication, while significantly reducing computational costs, making it suitable for real-time applications in manufacturing, traffic, and logistics. Reinforcement Learning-Guided Beam Search (RLGBS) was proposed as a general framework for neural-guided search in exponential search spaces with polynomial time budget. Unlike existing approaches, RLGBS relies solely on trained reinforcement learning (RL) policy for exploration at inference time and can operate on sparse reward definitions. A case study on a novel drying simulation environment modeled after a paper machine's multi-cylinder dryer section demonstrated energy savings over optimized baselines and generalization to unseen process operating conditions. Building upon RLGBS, Reinforcement Learning-Guided Constrained Beam Search (RLCBS) was developed to incorporate design constraints into RL policy-generated action sequences flexibly at inference time. The method supports unique positive constraints not studied by the existing literature that can be used to force desired actions into the generated sequence in the way that maximizes overall RL-rated sequence probability. Experiments on a novel modular Smart Dryer test bed demonstrated RLCBS's effectiveness in optimizing process configurations under multiple inference-time constraints, outperforming NSGA-II across various operating conditions with significant speed improvements.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Chen, Siyuan

Contributors dc:contributor

• Shao, Chenhui
• Salapaka, Srinivasa M
• Ferreira, Placid M
• Wang, Pingfeng

• Radiofrequency Identification
• Multi-object Tracking
• Tracking-by-detection
• Online Tracking
• Transportation
• Reinforcement Learning
• Beam Search
• Drying
• Papermaking
• Process Optimization
• Decarbonization
• Constrained Beam Search
• Rotors
• Fault Diagnosis
• Neural Networks
• Convolution
• Kernel
• Vibrations
• Condition Monitoring
• Rotating Machinery
• Rotor And Bearing Systems
• Kalman Filters
• Detectors
• Visualization
• Task Analysis
• Reliability
• Real-time Systems

Statement dc:rights

• Copyright 2024 Siyuan Chen

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/127244

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F127244