Virtual Sensor Data Imputation Using Generative Adversarial Imputation Nets and Pearson Correlation

The term Internet of things (IoT) was introduced into every areas of life; more and more devices are connecting and operating that rely on information collected by IoT sensors. As a matter of fact, sensor failures may either kill the operation of a system or make it malfunctioning because of the interrupted flow of data. Consequently, missing data is a pervasive problem with real-world datasets. Besides that, machine learning methods have recently been applied in several applications and achieved state-of-the art results. Therefore, this work leverages machine learning methods to resolve the missing-data problem. This paper proposes a novel method named PGAIN Virtual Sensor (PGAIN-VS) based on virtual sensors to impute missing data. Remarkably, PGAIN-VS uses the Pearson metric to evaluate the correlations of data among data-collected devices to remove noise and bias potentially causing inaccurate data imputation. PGAIN-VS can produce good estimates for the missing values, so it is absolutely capable of replacing physical sensors during failure time. We evaluated our method on several datasets and compared it to other recent works. Furthermore, the experiments showed that our approach achieved better performance up to 20% in the considered datasets with different metrics.