Imputation of Missing Values in Time Series Using an Adaptive-Learned Median-Filled Deep Autoencoder

Area of Science:

• Data Science
• Machine Learning
• Industrial Engineering

Background:

• Missing values are common in industrial datasets due to sensor faults and transmission issues.
• Incomplete data hinders data utilization and degrades data-driven model performance.
• Existing imputation methods struggle with large amounts of missing data and often ignore time-series autocorrelation.

Purpose of the Study:

• To propose an unsupervised method for imputing missing values in industrial time-series data.
• To address the limitations of supervised learning methods when dealing with extensive missing data.
• To account for the autocorrelation inherent in time-series data.

Main Methods:

• An adaptive-learned median-filled deep autoencoder (AM-DAE) was developed.
• The AM-DAE uses an unsupervised approach, continuously updating missing values with medians from input and reconstruction.
• An adaptive learning strategy focuses reconstruction on missing or non-missing values during training.

Main Results:

• The AM-DAE demonstrated superior performance in imputing missing values in industrial time-series data.
• The method effectively handles datasets with a significant proportion of missing values.
• Validation on two industrial examples confirmed the proposed method's effectiveness compared to advanced techniques.

Conclusions:

• The proposed AM-DAE offers an effective unsupervised solution for missing value imputation in industrial time-series data.
• The adaptive learning strategy enhances the imputation accuracy by focusing on critical data points.
• This approach improves the reliability and performance of data-driven models in industrial applications.