The Use of Supervised Machine Learning for Classification Purposes with a Method for Sensor Reduction

Abstract

Rapid adoption of sensor-based feedback and control systems in smart gadgets. Markets that place a premium on affordability are among the most likely to embrace these devices. Conventional machine learning-based control systems often incorporate data from several sensors in order to achieve performance objectives. Another method is presented that uses the time series data collected by a single sensor. Domain experts' knowledge of the system's physical occurrences is used to segment the time series output into discrete time chunks. The machine learning system's characteristics are derived from statistical observations over many time periods. When more characteristics are found that decouple vital physical measurements, the system's performance is improved. This state-of-the-art approach requires fewer observations than conventional methods, yet produces equivalent precision. Because of the reduced number of sensors and the considerably streamlined and more robust algorithm development and testing stage, the resulting development effort is far more cost-effective than that of traditional sensor categorization systems. The authors present their conclusions by analyzing a case study of a media-type classification system used in a commercially available printing system.