A control charts methodology to multistage process monitoring and root cause diagnosis using hierarchical Bayesian networks

Mondal, Partha Protim

Permalink

https://hdl.handle.net/2142/121925 Copy

Description

Title

A control charts methodology to multistage process monitoring and root cause diagnosis using hierarchical Bayesian networks

Author(s)

Mondal, Partha Protim

Issue Date

2023-07-31

Director of Research (if dissertation) or Advisor (if thesis)

• Kapoor, Shiv G
• Ferreira, Placid M

Doctoral Committee Chair(s)

• Kapoor, Shiv G
• Ferreira, Placid M

Committee Member(s)

• Shao, Chenhui
• Sreenivas, Ramavarapu S
• Bless, Patrick N

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Multistage manufacturing systems
• Bayesian networks
• Quality control: Control charts
• Process monitoring
• Root cause diagnosis

Abstract

Multistage manufacturing systems (MMS) are ubiquitous in the production of modern products that are used in today’s world. Examples include semiconductor manufacturing and packaging, automotive manufacturing and assembly, pharmaceutical manufacturing, among others. An MMS involves sequential stages that collectively produce a final product. In an MMS, there are many interdependent stages with complex interactions between them. Furthermore, any process variations that occur in one stage of an MMS can propagate to all subsequent stages, leading to a buildup of variation that affects the quality of the final product. This makes it challenging to pinpoint specific source of any quality related issues. Although control charts are the most widely used tool for quality monitoring in an MMS, they do not provide insights into the root causes, necessitating additional diagnostic actions. Conventional root cause diagnosis methods typically rely on either accurate input-output models or comprehensive fault pattern libraries, both of which are difficult to construct and often not available. However, the digitalization of manufacturing, including Industrial Internet of Things (IIoT) and cloud-based computing services to collate, curate and store large amounts of process data along with modern data science approaches, alleviate some of these difficulties by enabling automated shop-floor smart decision making. In light of these considerations, it becomes crucial to develop and evaluate data-driven methodologies to address analytically challenging manufacturing problems such as the root cause diagnosis problem of observed quality issues in an MMS. In this research work, using Bayesian networks, a novel Absolute Mean Deviation of States (AMDS) control charting methodology has been developed for multistage process monitoring and root cause diagnosis. The methodology uses trained Bayesian networks to predict/infer the state of each input in a multistage process using observed/measured responses. Subsequently, the stream of inferences, consisting of predicted input states, is utilized to construct AMDS control charts to track the current state of each input in the multistage process to identify if any input has deviated from its set distribution, causing the process to go out-of-control. This enables simultaneous detection of out-of-control processes and diagnosis of root causes. The developed methodology is thoroughly tested using randomly generated non-linear 2-stage processes. This involve simulating out-of-control scenarios by deliberately altering the mean or the variance of the multistage process inputs from their predefined distribution. The results show consistent performance in identifying out-of-control processes and accurately diagnosing the underlying causes. Investigative simulations further confirm the methodology’s good performance in terms of average run length (ARL) and detection power (DP) performance metrics. Moreover, the methodology is validated using a custom-made two-stage micro hot embossing process, successfully diagnosing root causes in various process mean shift fault scenarios, including a dual faults scenario. Bayesian network being a crucial component in the AMDS control charts methodology, a framework has been developed to integrate manufacturing knowledge sources, such as Failure Mode and Effect Analysis (FMEA) and hierarchical variable constraints, with data-driven structure learning algorithms to enhance the accuracy and reliability of the Bayesian network model. Furthermore, the framework provides a way to sequentially learn Bayesian network structure for large multistage processes, making the learning process flexible and tractable. Additionally, comparative simulations are conducted to investigate the effectiveness of the knowledge sources integrated networks. The results reveal that the knowledge source integrated networks exhibit superior performance compared to solely data-driven networks in detecting process faults caused by mean shifts or variance changes in the process inputs and diagnosing the corresponding root causes. This superior performance has been observed across diverse training conditions, encompassing scenarios involving 2-, 4-, and 8-stage processes. Considering that the Bayesian network predicted discrete input states typically follow a multinomial distribution, therefore, using the multinomial distribution, Q-control charting methodology has been developed, where the Q-statistic refers to the count of a specific discrete state occurring in a predefined sample size. The Q-control chart offers a statistically grounded alternative to the heuristically derived AMDS control chart. Notably, the control limits for the Q-control chart can be pre-established using multinomial probability distribution, whereas the AMDS control chart relies on iterative simulations to determine the control limits. Furthermore, to validate the Q-control chart, it is implemented in the two-stage micro hot embossing process. Results show that the Q-control charts are able to successfully detect out-of-control processes and accurately diagnose the root causes, which included mean alterations in critical process inputs such as setpoint pressure, heating temperature, and demolding temperature. In order to characterize and compare the performance of the AMDS and Q-control charts, comprehensive simulation studies are conducted across 2-, 4-, and 6-stage processes. The diverse testing conditions involve operating both the control charts with independent and running samples. However, as running samples are correlated, therefore, a Markov chain approach has been developed for the exact type-I error calculation of Q-control charts. The simulation results reveal that both control charts have comparable performance. Furthermore, as the complexity of the multistage process is increased by adding more stages, both control charts demonstrate good scalable performance. Additionally, when operated using running samples, both control charts showcase superior performance due to the increased resolution provided by running samples. To demonstrate the industrial applicability of Q-control charts, a practical implementation is conducted utilizing the Node-RED platform. Node-RED is a flow-based programming tool that enables rapid prototyping of applications. The implementation consists of a Node-RED workflow representing a virtual multistage process, a backend application for complex computations, and a dashboard application for visualizing the current state of the system. This integrated system is used to successfully simulate a running production line, showcasing real-time process monitoring and root cause diagnosis through graphical insights presented in the dashboard application.

Graduation Semester

2023-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2023 Partha Protim Mondal

Owning Collections