SUMMARY

Due to the recent movements in Industry 4.0 and Internet of Things (IoT), accessing or generating data in Smart Manufacturing (SM) domain has become more attainable; communication protocols such as MTConnect and OPC-UA provide access to a majority of raw data from machine tools while retrofit sensor packs facilitate high-frequency data acquisition from legacy and modern equipment. These technologies have led to the generation of quantities of raw data, known as Big Data (BD). Current IoT architectures and frameworks propose Cloud Computing (CC) and Centralized Training (CT) as the addressing solutions for BD and collaborative Machine Learning (ML) models. These solutions, however, have limitations such as Internet dependency and requiring expensive and high-performance cloud resources. Furthermore, by generating more data, a higher performance framework is required as cloud computing is applied to analyze larger datasets that are either historical in nature or generated from ever-increasing ubiquitous sensors and sensor arrays that are deployed in modern manufacturing operations. Studying IoT architectures and stream analytics as the foundations of creating successful IoT platforms is essential. In this regard, this study proposes a novel, high-performance, and data-driven IoT architecture that considers automated machine learning for manufacturing focusing on process control and a deeper understanding in manufacturing process and systems performance in the Cyber-Physical Systems (CPS) domain. First, a novel generalized 3-layer IoT architecture utilizing Edge Computing (EC), Fog Computing (FC), CC, and Federated Learning (FL) is presented, where data are preprocessed in the Edge layer, ML models are trained in the Fog layer and the resulting pieces of training are aggregated in a centralized cloud model. Second, two novel stream analytics engines of Outlier Detection and Bayesian Classification, capable of real-time (RT) incremental training and prediction are proposed and analyzed for this architecture. Results show that the training latency for both of the Outlier and Bayesian engines as well as the FL algorithms remained constant as the number of data points increased. On a 1000 datapoint dataset, the training performance for an upcoming datapoint for the Outlier and Bayesian engines were on average 136 and 48 times faster, respectively, than retraining the models with all of the data points. Results suggest that the methods discussed in the proposed architecture, therefore, can lead to the development of higher performance and more scalable IoT frameworks, which also require minimal storage and computing power.