Matching Items (21)
Description
Sequential event prediction or sequential pattern mining is a well-studied topic in the literature. There are a lot of real-world scenarios where the data is released sequentially. People believe that there exist repetitive patterns of event sequences so that the future events can be predicted. For example, many companies build their recommender system to predict the next possible product for the users according to their purchase history. The healthcare system discovers the relationships among patients’ sequential symptoms to mitigate the adverse effect of a treatment (drugs or surgery). Modern engineering systems like aviation/distributed computing/energy systems diagnosed failure event logs and took prompt actions to avoid disaster when a similar failure pattern occurs. In this dissertation, I specifically focus on building a scalable algorithm for event prediction and extraction in the aviation domain. Understanding the accident event is always the major concern of the safety issue in the aviation system. A flight accident is often caused by a sequence of failure events. Accurate modeling of the failure event sequence and how it leads to the final accident is important for aviation safety. This work aims to study the relationship of the failure event sequence and evaluate the risk of the final accident according to these failure events. There are three major challenges I am trying to deal with. (1) Modeling Sequential Events with Hierarchical Structure: I aim to improve the prediction accuracy by taking advantage of the multi-level or hierarchical representation of these rare events. Specifically, I proposed to build a sequential Encoder-Decoder framework with a hierarchical embedding representation of the events. (2) Lack of high-quality and consistent event log data: In order to acquire more accurate event data from aviation accident reports, I convert the problem into a multi-label classification. An attention-based Bidirectional Encoder Representations from Transformers model is developed to achieve good performance and interpretability. (3) Ontology-based event extraction: In order to extract detailed events, I proposed to solve the problem as a hierarchical classification task. I improve the model performance by incorporating event ontology. By solving these three challenges, I provide a framework to extract events from narrative reports and estimate the risk level of aviation accidents through event sequence modeling.
ContributorsZhao, Xinyu (Author) / Yan, Hao (Thesis advisor) / Liu, Yongming (Committee member) / Ju, Feng (Committee member) / Iquebal, Ashif (Committee member) / Arizona State University (Publisher)
Created2022
Description
Additive manufacturing consists of successive fabrication of materials layer upon layer to manufacture three-dimensional items. Several key problems such as poor quality of finished products and excessive operational costs are yet to be addressed before it becomes widely applicable in the industry. Retroactive/offline actions such as post-manufacturing inspections for defect detection in finished products are not only extremely expensive and ineffective but are also incapable of issuing corrective action signals during the building span. In-situ monitoring and optimal control methods, on the other hand, can provide viable alternatives to aid with the online detection of anomalies and control the process. Nevertheless, the complexity of process assumptions, unique structure of collected data, and high-frequency data acquisition rate severely deteriorates the performance of traditional and parametric control and process monitoring approaches. Out of diverse categories of additive manufacturing, Large-Scale Additive Manufacturing (LSAM) by material extrusion and Laser Powder Bed Fusion (LPBF) suffer the most due to their more advanced technologies and are therefore the subjects of study in this work. In LSAM, the geometry of large parts can impact the heat dissipation and lead to large thermal gradients between distance locations on the surface. The surface's temperature profile is captured by an infrared thermal camera and translated to a non-linear regression model to formulate the surface cooling dynamics. The surface temperature prediction methodology is then combined into an optimization model with probabilistic constraints for real-time layer time and material flow control. On-axis optical high-speed cameras can capture streams of melt pool images of laser-powder interaction in real-time during the process. Model-agnostic deep learning methods offer a great deal of flexibility when facing such unstructured big data and thus are appealing alternatives to their physical-related and regression-based modeling counterparts. A configuration of Convolutional Long-Short Term Memory (ConvLSTM) auto-encoder is proposed to learn a deep spatio-temporal representation from sequences of melt pool images collected from experimental builds. The unfolded bottleneck tensors are then further mined to construct a high accuracy and low false alarm rate anomaly detection and monitoring procedure.
ContributorsFathizadan, Sepehr (Author) / Ju, Feng (Thesis advisor) / Wu, Teresa (Committee member) / Lu, Yan (Committee member) / Iquebal, Ashif (Committee member) / Arizona State University (Publisher)
Created2022
Description
Recent advancements in machine learning methods have allowed companies to develop advanced computer vision aided production lines that take advantage of the raw and labeled data captured by high-definition cameras mounted at vantage points in their factory floor. We experiment with two different methods of developing one such system to automatically track key components on a production line. By tracking the state of these key components using object detection we can accurately determine and report production line metrics like part arrival and start/stop times for key factory processes. We began by collecting and labeling raw image data from the cameras overlooking the factory floor. Using that data we trained two dedicated object detection models. Our training utilized transfer learning to start from a Faster R-CNN ResNet model trained on Microsoft’s COCO dataset. The first model we developed is a binary classifier that detects the state of a single object while the second model is a multiclass classifier that detects the state of two distinct objects on the factory floor. Both models achieved over 95% classification and localization accuracy on our test datasets. Having two additional classes did not affect the classification or localization accuracy of the multiclass model compared to the binary model.
ContributorsPaulson, Hunter (Author) / Ju, Feng (Thesis director) / Balasubramanian, Ramkumar (Committee member) / Barrett, The Honors College (Contributor) / School of Mathematical and Statistical Sciences (Contributor) / Computer Science and Engineering Program (Contributor)
Created2022-05
Description
The use of Red Blood Cells (RBCs) is a pillar of modern health care. Annually, the lives of hundreds of thousands of patients are saved through ready access to safe, fresh, blood-type compatible RBCs. Worldwide, hospitals have the common goal to better utilize available blood units by maximizing patients served and reducing blood wastage. Managing blood is challenging because blood is perishable, its supply is stochastic and its demand pattern is highly uncertain. Additionally, RBCs are typed and patient compatibility is required.
This research focuses on improving blood inventory management at the hospital level. It explores the importance of hospital characteristics, such as demand rate and blood-type distribution in supply and demand, for improving RBC inventory management. Available inventory models make simplifying assumptions; they tend to be general and do not utilize available data that could improve blood delivery. This dissertation develops useful and realistic models that incorporate data characterizing the hospital inventory position, distribution of blood types of donors and the population being served.
The dissertation contributions can be grouped into three areas. First, simulations are used to characterize the benefits of demand forecasting. In addition to forecast accuracy, it shows that characteristics such as forecast horizon, the age of replenishment units, and the percentage of demand that is forecastable influence the benefits resulting from demand variability reduction.
Second, it develops Markov decision models for improved allocation policies under emergency conditions, where only the units on the shelf are available for dispensing. In this situation the RBC perishability has no impact due to the short timeline for decision making. Improved location-specific policies are demonstrated via simulation models for two emergency event types: mass casualty events and pandemic influenza.
Third, improved allocation policies under normal conditions are found using Markov decision models that incorporate temporal dynamics. In this case, hospitals receive replenishment and units age and outdate. The models are solved using Approximate Dynamic Programming with model-free approximate policy iteration, using machine learning algorithms to approximate value or policy functions. These are the first stock- and age-dependent allocation policies that engage substitution between blood type groups to improve inventory performance.
This research focuses on improving blood inventory management at the hospital level. It explores the importance of hospital characteristics, such as demand rate and blood-type distribution in supply and demand, for improving RBC inventory management. Available inventory models make simplifying assumptions; they tend to be general and do not utilize available data that could improve blood delivery. This dissertation develops useful and realistic models that incorporate data characterizing the hospital inventory position, distribution of blood types of donors and the population being served.
The dissertation contributions can be grouped into three areas. First, simulations are used to characterize the benefits of demand forecasting. In addition to forecast accuracy, it shows that characteristics such as forecast horizon, the age of replenishment units, and the percentage of demand that is forecastable influence the benefits resulting from demand variability reduction.
Second, it develops Markov decision models for improved allocation policies under emergency conditions, where only the units on the shelf are available for dispensing. In this situation the RBC perishability has no impact due to the short timeline for decision making. Improved location-specific policies are demonstrated via simulation models for two emergency event types: mass casualty events and pandemic influenza.
Third, improved allocation policies under normal conditions are found using Markov decision models that incorporate temporal dynamics. In this case, hospitals receive replenishment and units age and outdate. The models are solved using Approximate Dynamic Programming with model-free approximate policy iteration, using machine learning algorithms to approximate value or policy functions. These are the first stock- and age-dependent allocation policies that engage substitution between blood type groups to improve inventory performance.
ContributorsDumkrieger, Gina (Author) / Mirchandani, Pitu B. (Thesis advisor) / Fowler, John (Committee member) / Wu, Teresa (Committee member) / Ju, Feng (Committee member) / Arizona State University (Publisher)
Created2020
Description
Recent advances in manufacturing system, such as advanced embedded sensing, big data analytics and IoT and robotics, are promising a paradigm shift in the manufacturing industry towards smart manufacturing systems. Typically, real-time data is available in many industries, such as automotive, semiconductor, and food production, which can reflect the machine conditions and production system’s operation performance. However, a major research gap still exists in terms of how to utilize these real-time data information to evaluate and predict production system performance and to further facilitate timely decision making and production control on the factory floor. To tackle these challenges, this dissertation takes on an integrated analytical approach by hybridizing data analytics, stochastic modeling and decision making under uncertainty methodology to solve practical manufacturing problems.
Specifically, in this research, the machine degradation process is considered. It has been shown that machines working at different operating states may break down in different probabilistic manners. In addition, machines working in worse operating stage are more likely to fail, thus causing more frequent down period and reducing the system throughput. However, there is still a lack of analytical methods to quantify the potential impact of machine condition degradation on the overall system performance to facilitate operation decision making on the factory floor. To address these issues, this dissertation considers a serial production line with finite buffers and multiple machines following Markovian degradation process. An integrated model based on the aggregation method is built to quantify the overall system performance and its interactions with machine condition process. Moreover, system properties are investigated to analyze the influence of system parameters on system performance. In addition, three types of bottlenecks are defined and their corresponding indicators are derived to provide guidelines on improving system performance. These methods provide quantitative tools for modeling, analyzing, and improving manufacturing systems with the coupling between machine condition degradation and productivity given the real-time signals.
Specifically, in this research, the machine degradation process is considered. It has been shown that machines working at different operating states may break down in different probabilistic manners. In addition, machines working in worse operating stage are more likely to fail, thus causing more frequent down period and reducing the system throughput. However, there is still a lack of analytical methods to quantify the potential impact of machine condition degradation on the overall system performance to facilitate operation decision making on the factory floor. To address these issues, this dissertation considers a serial production line with finite buffers and multiple machines following Markovian degradation process. An integrated model based on the aggregation method is built to quantify the overall system performance and its interactions with machine condition process. Moreover, system properties are investigated to analyze the influence of system parameters on system performance. In addition, three types of bottlenecks are defined and their corresponding indicators are derived to provide guidelines on improving system performance. These methods provide quantitative tools for modeling, analyzing, and improving manufacturing systems with the coupling between machine condition degradation and productivity given the real-time signals.
ContributorsKang, Yunyi (Author) / Ju, Feng (Thesis advisor) / Pedrielli, Giulia (Committee member) / Wu, Teresa (Committee member) / Yan, Hao (Committee member) / Arizona State University (Publisher)
Created2020
Description
Modern manufacturing systems are part of a complex supply chain where customer preferences are constantly evolving. The rapidly evolving market demands manufacturing organizations to be increasingly agile and flexible. Medium term capacity planning for manufacturing systems employ queueing network models based on stationary demand assumptions. However, these stationary demand assumptions are not very practical for rapidly evolving supply chains. Nonstationary demand processes provide a reasonable framework to capture the time-varying nature of modern markets. The analysis of queues and queueing networks with time-varying parameters is mathematically intractable. In this dissertation, heuristics which draw upon existing steady state queueing results are proposed to provide computationally efficient approximations for dynamic multi-product manufacturing systems modeled as time-varying queueing networks with multiple customer classes (product types). This dissertation addresses the problem of performance evaluation of such manufacturing systems.
This dissertation considers the two key aspects of dynamic multi-product manufacturing systems - namely, performance evaluation and optimal server resource allocation. First, the performance evaluation of systems with infinite queueing room and a first-come first-serve service paradigm is considered. Second, systems with finite queueing room and priorities between product types are considered. Finally, the optimal server allocation problem is addressed in the context of dynamic multi-product manufacturing systems. The performance estimates developed in the earlier part of the dissertation are leveraged in a simulated annealing algorithm framework to obtain server resource allocations.
This dissertation considers the two key aspects of dynamic multi-product manufacturing systems - namely, performance evaluation and optimal server resource allocation. First, the performance evaluation of systems with infinite queueing room and a first-come first-serve service paradigm is considered. Second, systems with finite queueing room and priorities between product types are considered. Finally, the optimal server allocation problem is addressed in the context of dynamic multi-product manufacturing systems. The performance estimates developed in the earlier part of the dissertation are leveraged in a simulated annealing algorithm framework to obtain server resource allocations.
ContributorsJampani Hanumantha, Girish (Author) / Askin, Ronald (Thesis advisor) / Ju, Feng (Committee member) / Yan, Hao (Committee member) / Mirchandani, Pitu (Committee member) / Arizona State University (Publisher)
Created2020
Description
Degradation process, as a course of progressive deterioration, commonly exists on many engineering systems. Since most failure mechanisms of these systems can be traced to the underlying degradation process, utilizing degradation data for reliability prediction is much needed. In industries, accelerated degradation tests (ADTs) are widely used to obtain timely reliability information of the system under test. This dissertation develops methodologies for the ADT data modeling and analysis.
In the first part of this dissertation, ADT is introduced along with three major challenges in the ADT data analysis – modeling framework, inference method, and the need of analyzing multi-dimensional processes. To overcome these challenges, in the second part, a hierarchical approach, that leads to a nonlinear mixed-effects regression model, to modeling a univariate degradation process is developed. With this modeling framework, the issues of ignoring uncertainties in both data analysis and lifetime prediction, as presented by an International Standard Organization (ISO) standard, are resolved. In the third part, an approach to modeling a bivariate degradation process is addressed. It is developed using the copula theory that brings the benefits of both model flexibility and inference convenience. This approach is provided with an efficient Bayesian method for reliability evaluation. In the last part, an extension to a multivariate modeling framework is developed. Three fundamental copula classes are applied to model the complex dependence structure among correlated degradation processes. The advantages of the proposed modeling framework and the effect of ignoring tail dependence are demonstrated through simulation studies. The applications of the copula-based multivariate degradation models on both system reliability evaluation and remaining useful life prediction are provided.
In summary, this dissertation studies and explores the use of statistical methods in analyzing ADT data. All proposed methodologies are demonstrated by case studies.
In the first part of this dissertation, ADT is introduced along with three major challenges in the ADT data analysis – modeling framework, inference method, and the need of analyzing multi-dimensional processes. To overcome these challenges, in the second part, a hierarchical approach, that leads to a nonlinear mixed-effects regression model, to modeling a univariate degradation process is developed. With this modeling framework, the issues of ignoring uncertainties in both data analysis and lifetime prediction, as presented by an International Standard Organization (ISO) standard, are resolved. In the third part, an approach to modeling a bivariate degradation process is addressed. It is developed using the copula theory that brings the benefits of both model flexibility and inference convenience. This approach is provided with an efficient Bayesian method for reliability evaluation. In the last part, an extension to a multivariate modeling framework is developed. Three fundamental copula classes are applied to model the complex dependence structure among correlated degradation processes. The advantages of the proposed modeling framework and the effect of ignoring tail dependence are demonstrated through simulation studies. The applications of the copula-based multivariate degradation models on both system reliability evaluation and remaining useful life prediction are provided.
In summary, this dissertation studies and explores the use of statistical methods in analyzing ADT data. All proposed methodologies are demonstrated by case studies.
ContributorsFANG, GUANQI (Author) / Pan, Rong (Thesis advisor) / Montgomery, Douglas C. (Committee member) / Ju, Feng (Committee member) / Hong, Yili (Committee member) / Arizona State University (Publisher)
Created2020
Description
Drinking water quality violations are widespread in the United States and elsewhere in the world. More than half of Americans are not confident in the safety of their tap water, especially after the 2014 Flint, Michigan water crisis. Other than accidental contamination events, stagnation is a major cause of water quality degradation. Thus, there is a pressing need to build a real-time control system that can make control decisions quickly and proactively so that the quality of water can be maintained at all times. However, towards this end, modeling the dynamics of water distribution systems are very challenging due to the complex fluid dynamics and chemical reactions in the system. This challenge needs to be addressed before moving on to modeling the optimal control problem. The research in this dissertation leverages statistical machine learning approaches in approximating the complex water system dynamics and then develops different optimization models for proactive and real-time water quality control. This research focuses on two effective ways to maintain water quality, flushing of taps and injection of chlorine or other disinfectants; both of these actions decrease the equivalent “water age”, a useful proxy for water quality related to bacteria growth. This research first develops linear predictive models for water quality and subsequently linear programming optimization models for proactive water age control via flushing. The second part of the research considers both flushing and disinfectant injections in the control problem and develops mixed integer quadratically constrained optimization models for controlling water age. Different control strategies for disinfectant injections are also evaluated: binary on-off injections and continuous injections. In the third part of the research, water demand is assumed to be uncertain and stochastic. The developed approach to control the system relates to learning the optimal real-time flushing decisions by combing reinforced temporal-difference learning approaches with linear value function approximation for solving approximately the underlying Markov decision processes. Computational results on widely used simulation models demonstrates the developed control systems were indeed effective for water quality control with known demands as well as when demands are uncertain and stochastic.
ContributorsLi, Xiushuang (Author) / Mirchandani, Pitu (Thesis advisor) / Boyer, Treavor (Committee member) / Ju, Feng (Committee member) / Pedrielli, Giulia (Committee member) / Arizona State University (Publisher)
Created2021
Description
Natural disasters are occurring increasingly around the world, causing significant economiclosses. To alleviate their adverse effect, it is crucial to plan what should be done in response
to them in a proactive manner. This research aims at developing proactive and real-time
recovery algorithms for large-scale power networks exposed to weather events considering
uncertainty. These algorithms support the recovery decisions to mitigate the disaster impact, resulting in faster recovery of the network. The challenges associated with developing
these algorithms are summarized below:
1. Even ignoring uncertainty, when operating cost of the network is considered the problem will be a bi-level optimization which is NP-hard.
2. To meet the requirement for real-time decision making under uncertainty, the problem could be formulated a Stochastic Dynamic Program with the aim to minimize
the total cost. However, considering the operating cost of the network violates the
underlying assumptions of this approach.
3. Stochastic Dynamic Programming approach is also not applicable to realistic problem sizes, due to the curse of dimensionality.
4. Uncertainty-based approaches for failure modeling, rely on point-generation of failures and ignore the network structure.
To deal with the first challenge, in chapter 2, a heuristic solution framework is proposed, and its performance is evaluated by conducting numerical experiments. To address the second challenge, in chapter 3, after formulating the problem as a Stochastic Dynamic Program, an approximated dynamic programming heuristic is proposed to solve the problem. Numerical experiments on synthetic and realistic test-beds, show the satisfactory performance of the proposed approach. To address the third challenge, in chapter 4, an efficient base heuristic policy and an aggregation scheme in the action space is proposed. Numerical experiments on a realistic test-bed verify the ability of the proposed method to
recover the network more efficiently. Finally, to address the fourth challenge, in chapter 5, a simulation-based model is proposed that using historical data and accounting for the interaction between network components, allows for analyzing the impact of adverse events on regional service level. A realistic case study is then conducted to showcase the applicability of the approach.
ContributorsInanlouganji, Alireza (Author) / Pedrielli, Giulia (Thesis advisor) / Mirchandani, Pitu (Committee member) / Reddy, T. Agami (Committee member) / Ju, Feng (Committee member) / Arizona State University (Publisher)
Created2021
Description
A production system is commonly restricted by time windows. For example, perishability is a major concern in food processing and requires products, such as yogurt, beer and meat, not to stay in buffer for long. Semiconductor manufacturing is faced with oxidation and moisture absorption issues, if a product in buffer is exposed to air for long. Machine reliability is a major source of uncertainty in production systems that causes residence time constraints to be unsatisfied, leading to potential product quality issues. Rapid advances in sensor technology and automation provide potentials to manage production in real time, but the system complexity, brought by residence time constraints, makes it difficult to optimize system performance while providing a guaranteed product quality. To contribute to this end, this dissertation is dedicated to modeling, analysis and control of production systems with constrained time windows. This study starts with a small-scale serial production line with two machines and one buffer. Even the simplest serial lines could have too large state space due to the consideration of residence time constraints. A Markov chain model is developed to approximately analyze its transient behavior with a high accuracy. An iterative learning algorithm is proposed to perform real-time control. The analysis of two-machine serial line contributes to the further analysis of more general and complex serial lines with multiple machines. Residence time constraints can be required in multiple stages. To deal with it, a two-machine-one-buffer subsystem isolated from a multi-stage serial production line is firstly analyzed and then acts as a building block to support the aggregation method for overall system performance. The proposed aggregation method substantially reduces the complexity of the problem while maintaining a high accuracy. A decomposition-based control approach is proposed to control a multi-stage serial production line. A production system is decomposed into small-scale subsystems, and an iterative aggregation procedure is then used to generate a production control policy. The decomposition-based control approach outperforms general-purpose reinforcement learning method by delivering significant system performance improvement and substantial reduction on computation overhead. Semiconductor assembly line is a typical production system, where products are restricted by time windows and production can be disrupted by machine failures. A production control problem of semiconductor assembly line is presented and studied, and thus total lot delay time and residence time constraint violation are minimized.
ContributorsWang, Feifan (Author) / Ju, Feng (Thesis advisor) / Askin, Ronald (Committee member) / Mirchandani, Pitu (Committee member) / Patel, Nital (Committee member) / Arizona State University (Publisher)
Created2021