Prahalada Rao

In-process Monitoring of Material Cross-Contamination Defects in Laser Powder Bed Fusion

The goal of this work is to detect the onset of material cross-contamination in laser powder bed fusion (L-PBF) additive manufacturing (AM) process using data from in situ sensors. Material cross-contamination refers to trace foreign materials that may be introduced in the powder feedstock used in the process due to such reasons as, poor cleaning of the machine after previous builds, or inadequate quality control during production and storage of the powder. Material crosscontamination may lead to deleterious changes in the microstructure of the AM part and consequently affect its functional properties. Accordingly, the objective of this work is to develop and apply a spectral graph theoretic approach to detect the occurrence of material crosscontamination in real-time as the part is being built using in-process sensors. The central hypothesis is that transforming the process signals in the spectral graph domain leads to early and more accurate detection of material cross-contamination in L-PBF compared to the traditional delay-embedded Bon-Jenkins stochastic time series analysis techniques, such as autoregressive (AR) and autoregressive moving average (ARMA) modeling. To test this hypothesis, Inconel alloy 625 (UNS alloy 06625) test parts were made at Edison Welding Institute (EWI) on a custom-built L-PBF apparatus integrated with multiple sensors, including a silicon photodetector (with 300 nm to 1100 nm optical wavelength). During the process two types of foreign contaminant materials, namely, tungsten and aluminum particulates under varying degrees of severity were introduced. To detect cross-contamination in the part, the photodetector sensor signatures were monitored hatch-by-hatch in the form of spectral graph transform coefficients. These spectral graph coefficients are subsequently tracked on a Hotelling T statistical control chart. Instances of Type II statistical error, i.e., probability of failing to detect the onset of material cross-contamination, was verified against X-ray computed tomography (XCT) scans of the part to be within 5% in the case of aluminum contaminant particles. In contrast, traditional stochastic time series modeling approaches, e.g., ARMA had corresponding Type II error exceeding 15%. Furthermore, the computation time for the spectral graph approach was found to be less than one millisecond, compared to nearly 100 milliseconds for the traditional time series models tested.

Layer-wise spatial modeling of porosity in additive manufacturing

Abstract The objective of this work is to model and quantify the layer-wise spatial evolution of porosity in parts made using Additive Manufacturing (AM) processes. This is an important research area because porosity has a direct impact on the functional integrity of AM parts such as their fatigue life and strength. To realize this objective, an Augmented Layer-wise Spatial log Gaussian Cox process (ALS-LGCP) model is proposed. The ALS-LGCP approach quantifies the spatial distribution of pores within each layer of the AM part and tracks their sequential evolution across layers. Capturing the layer-wise spatial behavior of porosity leads to a deeper understanding of where (at what location), when (at which layer), and to what severity (size and number) pores are formed. This work therefore provides a mathematical framework for identifying specific pore-prone areas in an AM part, and tracking the evolution of porosity in AM parts in a layer-wise manner. This knowledge is essential for initiating remedial corrective actions to avoid porosity in future parts, e.g., by changing the process parameters or part design. The ALS-LGCP approach proposed herein is a significant improvement over the current scalar metric used to quantify porosity, namely, the percentage porosity relative to the bulk part volume. In this article, the ALS-LGCP approach is tested for metal parts made using a binder jetting AM process to model the layer-wise spatial behavior of porosity. Based on offline, non-destructive X-Ray computed tomography (XCT) scan data of the part the approach identifies those areas with high risk of porosity with statistical fidelity approaching 85% (F-score). While the proposed work uses offline XCT data, it takes the critical first-step from a data analytics perspective for taking advantage of the recently reported breakthroughs in online, in-situ X-Ray-based monitoring of AM processes. Further, the ALS-LGCP approach is readily extensible for porosity analysis in other AM processes; our future forays will focus on improving the computational tractability of the approach for online monitoring.

Graph Theoretic Compressive Sensing Approach for Classification of Global Neurophysiological States from Electroencephalography (EEG) Signals

We present a data fusion framework integrating graph theoretic and compressive sensing (CS) techniques to detect global neurophysiological states using high-resolution electroencephalography (EEG) recordings. Acute stress induction (and control procedures) were used to elicit distinct states of neurophysiological arousal. We recorded EEG signals (128 channels) from 50 participants under two different states: hand immersion in room temperature water (control condition) or in chilled (~3 °C) water (stress condition). Thereafter, spectral graph theoretic Laplacian eigenvalues were extracted from these high-resolution EEG signals. Subsequently, the CS technique was applied for the classification of acute stress using the Laplacian eigenvalues as features. The proposed method was compared to a support vector machine (SVM) approach using conventional statistical features as inputs. Our results revealed that the proposed graph theoretic compressive sensing approach yielded better classification performance (~90 % F-score) compared to SVM with statistical features (~50 % F-Score). This finding indicates that the spectral graph theoretic compressive sensing approach presented in this work is capable of classifying global neurophysiological arousal with higher fidelity than conventional signal processing techniques.

Acute Stress Detection Using Recurrence Quantification Analysis of Electroencephalogram (EEG) Signals

In the present work we intend to classify the brain states under physical stress and experimental control conditions based on the nonlinear features of electroencephalogram (EEG) dynamics using support vector machine (SVM) and least absolute shrinkage and selection operator (LASSO). Recurrence Quantification Analysis (RQA) method was employed to quantify the nonlinear features of high-density electroencephalogram (EEG) signals recorded either during instances of acute stress induction or comparison conditions. Four RQA measures, including determinism (DET), entropy (ENTR), laminarity (LAM) and trapping time (TT) were extracted from the EEG signals to characterize the deterministic features of cortical activity. Results revealed that LASSO was highly efficient in classifying the conditions using any one of the selected RQA measures, while SVM achieved accurate classification based solely on ENTR and TT. Among all four measures of non-linear dynamics, ENTR yielded the best overall classification accuracy.