Self-Supervised Deep Hadamard Autoencoders for Treating Missing Data: A Case Study in Manufacturing

Data collected from sensors are pivotal to the Industrial Internet of Things (IIoT) applications as they would not be usable if the data quality is bad. Missing values commonly appear in industrial datasets due to several reasons. In particular, unexpected changes in the process, failure of the sensors that are recording the measurements due to network connectivity or hardware issues, censored/anonymous features and human error are some of the reasons for missing values in these datasets. In the manufacturing context, it is expensive and time consuming to collect all the different types of data related to a part leading to siloed and incomplete datasets. Due to characteristics such as presence of high missingness ratios, varying patterns of missingness, class imbalance and small-sized datasets, approaches such as single imputation or dropping the rows with missing values as is done in complete case analysis would result in loss of information and can also introduce a bias in the analysis depending on the underlying mechanism of missingness. The approach used for treatment of missing values must be carefully chosen based on the characteristics of the data at hand as the downstream generalization performance of algorithms trained on the imputed data would directly depend on the quality of imputation. In this work, we present a novel approach to treating missing values in a real-world dataset. Our approach uses deep Hadamard autoencoders (HAE) in addition to the self-supervised learning paradigm and shows better imputation performance on the unobserved entries of the test set as compared to approaches such as single imputation using mean, or other methods such as standard autoencoders (SAE). We demonstrate the effectiveness of the proposed method using a case study on a high-pressure die casting dataset (HPDC).