CM-GAN: A Cross-Modal Generative Adversarial Network for Imputing Completely Missing Data in Digital Industry.

Area of Science:

• Digital industry
• Data science
• Machine learning

Background:

• Multimodal data fusion is crucial for environmental awareness in digital industries.
• Missing time-series data, including complete data missing (CDM), hinders accurate modeling due to communication failures or cyberattacks.
• Existing imputation models fail to address CDM, where units are unobservable for extended periods.

Purpose of the Study:

• To propose a novel method capable of imputing missing data in scenarios of complete data missing (CDM).
• To develop a model that can generate long-term time-series data from existing spatio-temporal data for imputation.
• To enhance environmental awareness and improve prediction accuracy in digital industrial systems.

Main Methods:

• A novel cross-modal generative adversarial network (CM-GAN) was developed, integrating cross-modal data fusion and deep adversarial generation.
• The CM-GAN constructs a cross-modal data generator to produce synthetic long-term time-series data.
• Imputation is performed by replacing missing values with the generated data.

Main Results:

• Experiments on photovoltaic (PV) power output datasets demonstrated CM-GAN's superior performance over baseline models, achieving state-of-the-art results.
• Ablation studies confirmed the contribution of cross-modal data fusion and validated parameter settings.
• PV data imputed by CM-GAN improved the predictability and accuracy of deep learning prediction models.

Conclusions:

• CM-GAN effectively addresses the challenge of complete data missing (CDM) in industrial time-series data.
• The proposed method offers a robust solution for data imputation, enhancing environmental awareness and predictive capabilities.
• Generated data by CM-GAN provides valuable information for improving deep learning-based prediction models in industrial applications.