Novel data reductions techniques for predictive maintenance in the Internet of Things

There has been significant growth in adaptation of the Internet of Things (IoT) in various applications lately. Some common examples of IoT applications are intelligent transportation systems, smart healthcare, weather prediction, smart agriculture, and innovative production environments. In all these applications, sensor nodes are used that are usually tiny but can sense physical phenomena such as temperature, pressure, and sound, to name a few. These nodes generate a massive amount of data that needs to be transmitted, stored, and processed for insights to make decisions. For example, temperature or pressure data can be collected in intelligent production environments. Thanks to recent advancements in the area of Machine and Deep Learning, collected data can be used to train machine/deep learning models, and as a result, models can make predictions on unseen data. However, due to the resources required to transmit, store, process, and train ML models on massive data, efforts are being made to reduce the data size since data are used by ML models that predict physical phenomena(for example, if a machine will fail in given number of days or not), it is clear that data reduction needs to be done carefully. In this thesis, we develop and investigate data reduction approaches for IoT. We make two arguments: (a) Before developing data reduction techniques, it is vital to explore \textit{where} in an IoT network it should occur and (b) the data reduction approach could leverage the information that is already there. On top of traditional components (sensor and cloud), there is another component called EC. The EC promises computational capabilities near the data source. Therefore, the \textit{Sensor-Cloud} continuum is transformed into \textit{Sensor-Edge-Cloud}. Thus, resources offered by EC can be leveraged for data reduction in an IoT network. In this thesis, we formulate research questions that are answered by novel contributions. More specifically, we make the following contributions to this thesis: 1) We unfold \textit{where} perspective of data reduction approaches in the \textit{Sensor-Edge-Cloud}. 2) We propose novel algorithms that consider the information already there: variation in the data, the difference between values, and the accuracy of the downstream prediction model. The proposed algorithms promise significant data reduction without compromising the accuracy of the downstream prediction mode. These contributions will significantly help in advancements of data reduction approaches in IoT that will eventually help efficient machine learning (using reduced data)., thereby saving resources of the IoT network.