Big Data Analytics and Its Applications in Soft Sensor for Smart Manufacturing
Type of DegreePhD Dissertation
MetadataShow full item record
This dissertation presents research performed to develop industrial internet of things enabled testbed and statistics pattern analysis feature-based modelling approach for smart manufacturing. Advent of internet completely revolutionized the way humans communicate and share information. With rapid growth in internet users, amount of data generation and exchange increased exponentially resulting in ever increasing demand for faster internet. This demand drove multidisciplinary innovations resulting in fast & reliable wireless internet, cheap storage memory and efficient computation resources capable of processing large amount of data. Large amount of data generated by the users, when modeled effectively, offers the potential to make information exchange more efficient and productive. Internet of things is an extension of the same idea but with goal of improving efficiency, productivity and in turn profitability of a business. Network of things or items capable of exchanging information using some communication protocol is called Internet of things (IoT). Many health, retail and ecommerce businesses have benefited by modelling data generated from these IIoT devices. Combination of IIoT sensors and data modelling using machine learning techniques presents manufacturing industry with immense potential to identify and mitigate bottlenecks resulting into a more efficient manufacturing or smart-manufacturing. Smart-manufacturing is still at its incipient and much more research is required both at academic and industrial scales. In this work, Industrial internet of things enabled test-bed was developed to demonstrate entire smart-manufacturing project pipeline. It also provides idea about how a normal unit operations can be transformed into smart unit operation which is capable of real time process monitoring. This study provides discussion on sensor selection, micro-controller selection, data transfer and storage architecture, data compilation, noise characteristics over type of internet connections, data mining, possible way of dealing with data synchronization, model based data filtering by introducing new statistical approach. Moreover comparison study of recurrent neural networks, artificial neural networks and hierarchical modelling approach for prediction of flowrate and pump RPM will be presented. Another important aspect of smart manufacturing that was addressed is development of modelling approaches that can capture relevant information from a complex system while making resulting models more robust, simpler and accurate. A new feature based modelling approach has been proposed for spectrum dataset. In this approach relevant features are engineered from original raw spectrum with the help of different statistics. The proposed approach is compared with different conventional soft sensor approaches. Moreover this comparison study was carried out on four publically available industrial datasets. A Monte-Carlo validation and testing procedure was also devised in order to clearly indicate identify if any modelling approach’s performance is better or it just seems better because of bias variance tradeoff.