Understanding Manufacturing Process Dynamics and Machine Tool Anomaly Detection through Process Sensing and Machine Learning

As one crucial element in smart manufacturing, advanced sensing-enabled manufacturing equipment condition monitoring and anomaly detection are beneficial for predictive maintenance and reliability assurance. Widely adopted anomaly detection techniques, such as proximity-based comparison of compare sensing observations to benchmark references, have unguaranteed accuracy and reliability due to their limited abilities in distinguishing between the root causes for variations in observations: degradation of machine health conditions (performance anomaly occurrences) vs. changes in process and environmental conditions. Alternatively, machine health conditions can be essentially represented by the system input-output (process-observation) causal relationships, changes of which are sufficient and necessary conditions to imply anomaly occurrences. The difficulty that obstructs this approach is the causal discovery, due to the high-dimensionality, non-linearity, and non-stationarity associated with temporal modeling of system dynamics. The emerging deep learning (DL) techniques have demonstrated the ability to learn complex patterns and interrelations from multivariate data, thereby presenting a promising solution for quantitatively revealing the causal relationships of system observations against process and environmental inputs. Significant challenges that have remained are the physical interpretability of findings by DL models given their black-box nature, model complexity to ensure the real-time machine condition diagnosis, and the requirement on data amount and variety for model training that data need be acquired from both machine normal and abnormal conditions to ensure the model generalizability. The objective of this proposed research is to design an innovative and high-efficient DL architecture, with incremental learning capabilities from evolving data stream and physically interpretable decision makings, for the discovery of process-observation causal relationships in complex manufacturing equipment towards intelligent, robust, and real-time condition monitoring, anomaly detection, and root cause analysis. Intellectual Merit: The proposed research leverages recent advancement in process sensing and machine learning to advance the fundamental understanding of manufacturing process dynamics and equipment anomaly detection. Specifically, the research aims to differentiate the root causes for variations in sensing observations between equipment anomaly occurrences and changes in process/environmental conditions, through the process-observation causal relationships discovered by a DL model. To improve the trustworthiness of data-driven DL models in physics-based manufacturing shop floors and ensure the computational efficiency for real-time application purpose, an innovative DL architecture will be designed with capabilities on: 1) high-accuracy modeling and high-efficiency computation upon an optimal architecture without large manual tuning effort; 2) physically interpretable discovery of system input-output causal relationships; and 3) lessened requirement on training data amount and variety. The anticipated outcome and innovations include: 1) design a 1-Dimensional Convolutional Neural Network (1D-CNN) for sequential reasoning of process-observation-quality causal relationships, with optimal network architecture automatically searched; 2) endow physical interpretability of the 1D-CNN modeling through Reinforcement Learning-enabled integration of network training process with physical knowledge; 3) allow for modeling from unbalanced data and incremental learning for robust anomaly detection; and 4) provide high-quality manufacturing equipment fault simulation data to the community. Broader Impacts: McKinsey predicts machine learning-enhanced predictive maintenance will reduce annual maintenance cost by 10%, machine downtime by 20%, and inspection cost by 25% in manufacturing shop floors. This represents a winning profit margin for manufacturing industries, as maintenance cost typically accounts for 40-50% of a business’s operational budget. The challenge to achieve the goals comes from the lacking of applicable and trustable solutions that are specifically developed for manufacturing data analytics, which requires the analytics results to be consistent with domain knowledge and be interpretable by engineers. The proposed research addresses the challenge by synergistically integrating emerging DL techniques with physical domain knowledge to enable insights that are not available previously. If successful, the outcome will not only lead to reduced inspection time and maintenance cost, but also contribute to enhancing the science base for dynamical system modeling. Broad dissemination of the research outcome through publications and manufacturing conferences will impact a broad range of industry sectors in and beyond manufacturing. On the educational front, the proposed research will provide exciting opportunities for the participating graduate and undergraduate students by exposing them to truly interdisciplinary research that spans advanced manufacturing, process sensing, machine learning, and cloud computing, and significantly reinforce the existing curricula in manufacturing and data science at UK. The research outcomes will also be included into a senior undergraduate and graduate-level course at UK, process monitoring and machine learning, to have a new generation of mechanical and electrical engineers be knowledgeable at both physical science and data science. Active recruitment of women and under-represented groups in the project and outreach activities will further enhance the outcome dissemination and impact on the society at-large.