Title: Machine Learning Methods for Quality Monitoring in Manufacturing Applications
Date: Monday, October 19th, 2026
Time: 2:00 pm – 3:30 pm ET
Location: George Tower 1740
Microsoft Teams meeting
Join: https://teams.microsoft.com/meet/235126064810853?p=X0W0PNo3pdabWh1VnI
Meeting ID: 235 126 064 810 853
Passcode: H3rd2n9D
Committee:
Dr. Kamran Paynabar (co-Advisor), H. Milton Stewart School of Industrial and Systems
Engineering
Dr. Jianjun Shi (co-Advisor), H. Milton Stewart School of Industrial and Systems Engineering
Dr. Jing Li, H. Milton Stewart School of Industrial and Systems Engineering
Dr. Xiaochen Xian, H. Milton Stewart School of Industrial and Systems Engineering
Dr. Mostafa Gahrooei, Industrial & Systems Engineering at the University of Florida
Abstract:
Modern manufacturing systems are increasingly driven by complex, high-dimensional, and heterogeneous data collected from distributed production lines and advanced sensing technologies. These data-rich environments present new challenges for process modeling and monitoring, particularly in short-run and customized production settings where limited samples and variability across sites make traditional approaches ineffective. To address these challenges, this thesis develops a series of statistical and machine learning frameworks for high-dimensional process modeling and monitoring under practical constraints such as data heterogeneity, nonlinearity, and privacy preservation. Together these methodologies aim to advance the state of quality monitoring in modern manufacturing environments.
In chapter 2, a federated linear mixed-effects model is proposed to learn process parameters for short-run production settings. Short-run productions are common in manufacturing systems which require customization and flexibility under limited demand and resources. However, traditional statistical process control charts cannot handle such scenarios, since they either ask for data transformation or have strict assumptions which misalign with real-world applications. To address these limitations, we propose a federated framework to estimate process parameters across multiple manufacturing sites and create process monitoring baseline models for each line. Specifically, the restricted maximum likelihood method is used to estimate multivariate linear mixed model through a federated Expectation-Maximation (EM) algorithm. The fixed effect of the model captures the commonality amongst different manufacturing sites, while the random effect captures any heterogeneity between the sites. By using a federated setting to estimate the parameters, local data is kept at each of the manufacturing sites, allowing for increased data privacy and reducing the amount of bandwidth resources needed to transmit information between the local and global servers. Finally, we validate the model by using simulations and a case study using semi-generated paper mill manufacturing data.
Chapter 3 proposes two methods to detect changes in high-dimensional model relationships, also called profiles. Advancements in sensor and data collecting technology have resulted in data sets that are both high dimensional and heterogeneous, with forms such as scalars, waveform signals, images, videos, and 3D point clouds. These advancements have created the need to construct and monitor statistical models that are able to model the relationship between heterogeneous and high dimensional inputs and a high dimensional output. This chapter (1) discusses how to model the relationship between such inputs and output using Multiple Tensor-on-Tensor Regression (MTOT) and (2) presents different approaches for how to monitor this model relationship to detect potential changes over time. Monitoring the model relationship can be done either by monitoring the core tensor of the MTOT model and the residuals of new data samples. Both approaches retain the tensor structure of the profile – something that is lost if dimension reduction is done directly on the inputs and output data. This methodology is validated through several numerical studies including monitoring overlay error from semiconductor wafer shape data.
Finally, Chapter 4 proposes a robust nonlinear tensor on tensor regression model. To do this, a neural network is constructed that can capture the nonlinear relation between tensor inputs and a tensor output where the data is assumed to contain noise. The neural network architecture is an autoencoder with an encoder and decoder that uses tensor train layers (TT-layers) with rectified linear unit (ReLU) activations. The TT-layers enable large-scale tensor mapping with a reduced number of parameters while still preserving the tensor structure, while the ReLU activations capture the nonlinearity of the relationship between the network input and output tensors. Three different loss functions, Huber Loss, Log-Cosh Loss, and MSE Trimmed Loss, are evaluated through a series of numerical studies to explore performance across different noise data scenarios.