Conditional Diffusion Based Spatio-Temporal Imputation Model for Reliable and Intelligent Service in Consumer Electronics

With the significant progress of Artificial Intelligence and Internet of Things technologies, the demand for intelligent services and enhanced user experience in consumer electronics systems such as smart home devices and intelligent transportation terminals is rapidly increasing. The devices in this system generate massive spatio-temporal data that forms the foundation for reliable and intelligent service of consumer electronics. Inevitably, data collected by these devices is usually incomplete due to sensor failures, communication losses, or system malfunctions, which severely compromises the reliability of intelligent services and degrades user experience. To tackle this issue, researchers proposed imputation methods to fill in missing data. Among them, diffusion models have shown superiority for imputation tasks. However, existing diffusion based methods have some key limitations that affect practicality: long inference time due to iterative denoising steps, unrealistic masking strategies that ignore missing patterns in existing datasets, and overemphasis on irrelevant information in highly missing data segments. To overcome these challenges, this paper introduces a conditional diffusion based spatio-temporal imputation model for reliable and intelligent service (STIRS) in intelligent consumer electronic systems that balances both accuracy and efficiency in imputation tasks. Specifically, STIRS leverages an SDE solver to reduce the lengthy inference time and proposes a window-mixed masking strategy to better reflect the missing patterns in existing datasets. Additionally, integrity-based attention is employed to prioritize relevant information, thereby optimizing feature extraction for improved performance. Experiments on three real-world datasets with different missing data scenarios (10%-90%) demonstrate that STIRS achieves superior imputation performance while significantly reducing the inference time, is up to 4.3 × faster than the current state-of-the-art model.