The future of emerging smart manufacturing technologies is undoubtedly affected by the innovations in manufacturing processes and next-generation materials manufacture and characterization together with data science and AI. Towards this, my research presents innovations in two key areas:
Streaming data analytics forms the backbone of smart manufacturing systems to facilitate “real-time” decision-making.
For time series data, my research focused on harnessing the signal phase information (mostly neglected in the current literature) from noisy and transient streaming data via intrinsic time-scale decomposition and phase synchronization for fast and causal detection of process anomalies and change points.
For streaming image data, we developed an unsupervised image segmentation approach, with statistical consistency, by iteratively identifying the optimal graph cut and learning the parameters of the graph cut via maximum a posteriori estimation.
Most, if not all manufacturing processes involve some aspect of optimizing certain "objectives" of interest, e.g., increasing the material strength. Traditional solutions are either based on passive experimental design approaches or sometimes just on past experience. Accelerating the experimental process by adaptively updating the experimental design is the key to reducing the cycle time and material costs.
My research focuses on active learning models and approaches to integrate the knowledge of the process physics into enabling high fidelity search in the state space. On the methodological side, we will focus on the challenges emerging from the process uncertainty where it may not be feasible to guarantee strict improvements. Under such circumstances, how can we generate solutions that are correct with some “confidence”, i.e., Probably Approximately Correct. Applications include: identifying optimal process parameters in additive manufacturing, discovery of materials with novel properties, etc.
Quantification of process evolution is critical to realizing autonomous decision making in SM, yet only a handful of methods have addressed this problem. My research focuses on physics-based graph representations to capture the evolution in the morphology of manufacturing processes such as polishing and nanoparticle synthesis.
We employ a copula-based approach to learn the joint evolution of morphological features as gathered by images. Spectral characteristics of the resulting graphs showed significantly improved resolution (up to an order of magnitude) in terms of accurately resolving the process endpoints—deciding when to stop the process—as compared to conventionally employed surface roughness measurements. We also studied the robustness of the spectral characteristics under different graph perturbations. Via theoretical and numerical results, we established that the planar graphs achieve an average connectivity of six in the final stages of the polishing processes.