A Novel Data Transformation for Improving Predictive Accuracy of Online Missing Value Imputation During Reefer Container Monitoring

Internal ambiance monitoring of reefer containers using the Internet of Things (IoT) is important for cold chain tracking and traceability. However, disruptions due to power fluctuations, communication failures, and sensor malfunctions can lead to missing values in the data streams, impacting the effectiveness of the monitoring system. Existing algorithms addressing this issue assume the stationarity of the training dataset, ignoring the fact that the real-world datasets are mostly nonstationary. This article presents a two-stage framework to address this issue in the context of reefer container monitoring. The first stage detects the missing values and segments the training set, while the second stage employs a rotational transformation (RT), followed by an imputation algorithm. The framework is demonstrated using datasets collected during the remote monitoring of a reefer container, which is found to be nonstationary using the augmented Dickey-Fuller (ADF) test. The accuracy is evaluated using metrics such as RMSE, MAE, MAPE, and mean (bias) and shows better performance of the imputation algorithms with the proposed transformation compared to without it or with few other existing ones. Besides the imputation accuracy, the information preserved within the transformed dataset is evaluated using mutual information. The results suggest that the amount of retained information in the case of the proposed approach surpasses that of existing techniques. In addition, a metric to assess the predictive capability of an algorithm for a specific dataset is used to evaluate the framework, and the results indicate improved predictability compared to the existing ones.