Smart manufacturing is a revolutionary domain integrating advanced sensing technology, machine learning methods, and the industrial internet of things (IIoT). The development of sensing technology provides large amounts and various types of data (e.g., profile, image, point cloud, etc.) to describe each stage of a manufacturing process. The machine learning methods have the advantages of efficiently and effectively processing and fusing large-scale datasets and demonstrating outstanding performance in different tasks (e.g., diagnosis, monitoring, etc.). Despite the advantages of incorporating machine learning methods into smart manufacturing, there are some widely existing concerns in practice: (1) Most of the edge devices in the manufacturing system only have limited memory space and computational capacity; (2) Both the performance and interpretability of the data analytics method are desired; (3) The connection to the internet exposes the manufacturing system to cyberattacks, which decays the trustiness of data, models, and results. To address these limitations, this dissertation proposed systematic engineering-driven machine learning methods to improve the system intelligence for smart manufacturing. The contributions of this dissertation can be summarized in three aspects. First, tensor decomposition is incorporated to approximately compress the convolutional (Conv) layer in Deep Neural Network (DNN), and a novel layer is proposed accordingly. Compared with the Conv layer, the proposed layer significantly reduces the number of parameters and computational costs without decaying the performance. Second, a physics-informed stochastic surrogate model is proposed by incorporating the idea of building and solving differential equations into designing the stochastic process. The proposed method outperforms pure data-driven stochastic surrogates in recovering system patterns from noised data points and exploiting limited training samples to make accurate predictions and conduct uncertainty quantification. Third, a Wasserstein-based out-of-distribution detection (WOOD) framework is proposed to strengthen the DNN-based classifier with the ability to detect adversarial samples. The properties of the proposed framework have been thoroughly discussed. The statistical learning bound of the proposed loss function is theoretically investigated. The proposed framework is generally applicable to DNN-based classifiers and outperforms state-of-the-art benchmarks in identifying out-of-distribution samples. / Doctor of Philosophy / The global industries are experiencing the fourth industrial revolution, which is characterized by the use of advanced sensing technology, big data analytics, and the industrial internet of things (IIoT) to build a smart manufacturing system. The massive amount of data collected in the engineering process provides rich information to describe the complex physical phenomena in the manufacturing system. The big data analytics methods (e.g., machine learning, deep learning, etc.) are developed to exploit the collected data to complete specific tasks, such as checking the quality of the product, diagnosing the root cause of defects, etc. Given the outstanding performances of the big data analytics methods in these tasks, there are some concerns arising from the engineering practice, such as the limited available computational resources, the model's lack of interpretability, and the threat of hacking attacks. In this dissertation, we propose systematic engineering-driven machine learning methods to address or mitigate these widely existing concerns. First, the model compression technique is developed to reduce the number of parameters and computational complexity of the deep learning model to fit the limited available computational resources. Second, physics principles are incorporated into designing the regression method to improve its interpretability and enable it better explore the properties of the data collected from the manufacturing system. Third, the cyberattack detection method is developed to strengthen the smart manufacturing system with the ability to detect potential hacking threats.

Error equivalence concerns the mechanism whereby different error sources result in identical deviation and variation patterns on part features. This could have dual effects on process variation reduction: it significantly increases the complexity of root cause diagnosis in process control, and provides an opportunity to use one error source as based error to compensate the others. There are fruitful research accomplishments on establishing error equivalence methodology, such as error equivalence modeling, and an error compensating error strategy. However, no work has been done on developing an efficient process design approach by investigating error equivalence. Furthermore, besides the process mean shift, process fault also manifests itself as variation increase. In this regard, studying variation equivalence may help to improve the root cause identification approach. This thesis presents engineering driven approaches for process design and control via embedding error equivalence mechanisms to achieve a better, insightful understanding and control of manufacturing processes. The first issue to be studied is manufacturing process design and optimization based on the error equivalence. Using the error prediction model that transforms different types of errors to the equivalent amount of one base error, the research derives a novel process tolerance stackup model allowing tolerance synthesis to be conducted. Design of computer experiments is introduced to assist the process design optimization. Secondly, diagnosis of multiple variation sources under error equivalence is conducted. This allows for exploration and study of the possible equivalent variation patterns among multiple error sources and the construction of the library of equivalent covariance matrices. Based on the equivalent variation patterns library, this thesis presents an excitation-response path orientation approach to improve the process variation sources identification under variation equivalence. The results show that error equivalence mechanism can significantly reduce design space and release us from considerable symbol computation load, thus improve process design. Moreover, by studying the variation equivalence mechanism, we can improve the process diagnosis and root cause identification.

Models which describe the performance of physical process are essential for quality prediction, experimental planning, process control and optimization. Engineering models developed based on the underlying physics/mechanics of the process such as analytic models or finite element models are widely used to capture the deterministic trend of the process. However, there usually exists stochastic randomness in the system which may introduce the discrepancy between physics-based model predictions and observations in reality. Alternatively, statistical models can be used to develop models to obtain predictions purely based on the data generated from the process. However, such models tend to perform poorly when predictions are made away from the observed data points. This dissertation contributes to model enhancement research by integrating physics-based model and statistical model to mitigate the individual drawbacks and provide models with better accuracy by combining the strengths of both models. The proposed model enhancement methodologies including the following two streams: (1) data-driven enhancement approach and (2) engineering-driven enhancement approach. Through these efforts, more adequate models are obtained, which leads to better performance in system forecasting, process monitoring and decision optimization. Among different data-driven enhancement approaches, Gaussian Process (GP) model provides a powerful methodology for calibrating a physical model in the presence of model uncertainties. However, if the data contain systematic experimental errors, the GP model can lead to an unnecessarily complex adjustment of the physical model. In Chapter 2, we proposed a novel enhancement procedure, named as "Minimal Adjustment", which brings the physical model closer to the data by making minimal changes to it. This is achieved by approximating the GP model by a linear regression model and then applying a simultaneous variable selection of the model and experimental bias terms. Two real examples and simulations are presented to demonstrate the advantages of the proposed approach. Different from enhancing the model based on data-driven perspective, an alternative approach is to focus on adjusting the model by incorporating the additional domain or engineering knowledge when available. This often leads to models that are very simple and easy to interpret. The concepts of engineering-driven enhancement are carried out through two applications to demonstrate the proposed methodologies. In the first application where polymer composite quality is focused, nanoparticle dispersion has been identified as a crucial factor affecting the mechanical properties. Transmission Electron Microscopy (TEM) images are commonly used to represent nanoparticle dispersion without further quantifications on its characteristics. In Chapter 3, we developed the engineering-driven nonhomogeneous Poisson random field modeling strategy to characterize nanoparticle dispersion status of nanocomposite polymer, which quantitatively represents the nanomaterial quality presented through image data. The model parameters are estimated through the Bayesian MCMC technique to overcome the challenge of limited amount of accessible data due to the time consuming sampling schemes. The second application is to calibrate the engineering-driven force models of laser-assisted micro milling (LAMM) process statistically, which facilitates a systematic understanding and optimization of targeted processes. In Chapter 4, the force prediction interval has been derived by incorporating the variability in the runout parameters as well as the variability in the measured cutting forces. The experimental results indicate that the model predicts the cutting force profile with good accuracy using a 95% confidence interval. To conclude, this dissertation is the research drawing attention to model enhancement, which has considerable impacts on modeling, design, and optimization of various processes and systems. The fundamental methodologies of model enhancement are developed and further applied to various applications. These research activities developed engineering compliant models for adequate system predictions based on observational data with complex variable relationships and uncertainty, which facilitate process planning, monitoring, and real-time control.