A new approach to signal filtering method using K-means clustering and distance-based Kalman filtering

Abstract

Human axillary odours taken by an electronic nose (e-nose) device that uses a Metal Oxide Semiconductor (MOS) sensor not only contains a gas signal from the pure source of the axillary odour but also has the potential to contain other substances such as perfume and deodorant. This situation requires noise reduction so that dirty data can be cleaned and produce better predictions without wasting a lot of data. The approach taken in this study is to detect data clusters and data centroids from each reference data. Dimensional reduction using Linear Discriminant Analysis (LDA) on the reference data is carried out, then look for the centroid of each data using K-Means Clustering and use it to be a good signal estimate and process using Kalman Filtering so that it can be used to process axillary odour data containing deodorant. The proposed method was tested by a stacked Deep Neural Network (DNN) approach and can increase accuracy by 18.95% and balanced accuracy by 11.865% compared to original invalid data before filtering. The proposed method is also tested by other classification methods and able to produce the highest accuracy with 79.29% in Support Vector Classifier (SVC) and Multi-Layer Perception (MLP), while other filtering methods only get the highest accuracy with 69.03% also in SVC and MLP. We also analysed the execution time of each tested methods.