Leveraging Machine Learning Techniques to Obtain Data for Virtual Sensors

The Internet-of-Things (IoT) has revolutionised smart devices by enabling real-time monitoring through remote sensors. It is the most essential element particularly in smart sensing industrial applications such as environmental monitoring and industrial automation. These sensors provide crucial raw data to be analysed and accurate prediction of events of equipment breakdowns or preventive maintenance is required. However, if a physical sensor fails to function normally, virtual sensors can facilitate the missing data during downtime. Virtual sensors utilise predictive models to forecast the missing data, leveraging historical data and patterns from previously trained events to forecast sensor readings under the same conditions. In this research, the authors build a predictive model to generate data for a malfunctioned sensor by using actual data from other functional sensors. The hybrid setup between physical and virtual sensors will complement each other during operations to ensure fail-safe operation. In the research methodology, data from five sensors were analysed with predictive models of random forest. Data were trained on four of the sensors to predict the next day's readings of the fifth sensor. The experiment examined the impact of training various data durations (5, 10, and 15 days). The results revealed promising outcomes across all three training data sizes. Notably, the random forest regression model achieved better performance with larger training datasets, highlighting the impact of dataset size on model effectiveness.