Local Similarity Imputation Based on Fast Clustering for Incomplete Data in Cyber-Physical Systems

Missing values are common in cyber-physical systems (CPS) for a variety of reasons, such as sensor faults, communication malfunctions, environmental interferences, and human errors. An accurate missing value imputation is crucial to promote the data quality for data mining and statistical analysis tasks. Unfortunately, most of the existing methods take use of the whole data set to impute a missing value, which could have unfavorable influences and impacts (low accuracy or high complexity) on the imputed results caused by irrelevant records. Aiming at this problem, this paper develops a novel local similarity imputation method that estimates missing data based on fast clustering and top $k$-nearest neighbors. To improve the imputation accuracy, a two-layer stacked autoencoder combined with distinctive imputation is applied to locate the principal features of a dataset for clustering. Then, the top $k$ -nearest neighbor hybrid distance weighted imputation is approached to fill in missing values in clusters. The proposed method is evaluated on five popular University of California Irvine datasets as well as one air quality monitoring dataset collected from CPS through comparison with four high-quality existing imputation methods. Empirical results present that the proposed scheme can impute the missing data values effectively and efficiently, especially for the incomplete data with local characteristic in CPS.