SMOTE and Gaussian Noise Based Sensor Data Augmentation

Deep learning achieves successful prediction results by training multilayer neural network based machine learning models on large amounts of data. One of the best ways to improve performance of artificial neural networks is to add more data to the training set. In the literature, some data augmentation techniques have been developed for this purpose and they are widely used in processing image data. For example, data can be enriched by replicating the images in the training set with views from different angles and sufficient data can be obtained to find a generalizable model. In this study, we focused on augmenting sensor data by applying image data augmentation methods. A data set including temperature, humidity, light, and air quality sensor data was augmented using two different data augmentation techniques (SMOTE Regression and Gaussian Noise), and their effect on the performance of an LSTM model which estimates missing or incorrect values was investigated. RMSE values obtained by using real and augmented data were compared to evaluate the impact of both techniques. The most successful test estimation model from the data set was the air quality model. In addition, it was concluded that SMOTE regression gave better results when the two techniques were compared.