SeTFiN: Self-Data Generation and Temporal-Frequency Fusion for Multivariate Sensor Time-Series Forecasting

Abstract

Multivariate sensor time-series forecasting is important for reliable decision-making in industrial environments. However, sensor observations can be affected by degradation patterns, including spikes, noise, shifts, stuck values, and missing observations, which may reduce forecasting accuracy. To address this problem, this paper proposes SeTFiN, a Self-Data Generation and Temporal-Frequency Fusion Network for multivariate sensor time-series forecasting. SeTFiN integrates degradation-oriented Self-Data Generation (SDG) with temporal-frequency fusion. During training, SDG applies domain-specific perturbations at corresponding temporal and frequency patch indices without requiring additional labeled degraded data. The architecture performs frequency-guided Cross-Domain Feature Rectification (CDFR) before expert encoding and adaptively fuses temporal and frequency features after expert modeling. Experiments were conducted on five public forecasting benchmarks and the real-world industrial Semiconductor Fabrication Dataset (SFD). SeTFiN achieved competitive performance on the public benchmarks. In the three-seed clean-SFD evaluation, it achieved the lowest overall average errors among the compared models, with an MSE of 11.628 and an MAE of 0.870. Under five controlled synthetic degradation conditions applied to the SFD test inputs, SDG reduced MSE across all degradation types, while its effect on MAE varied. SeTFiN also remained competitive in several long-lookback settings, although its relative performance depended on the dataset and input length. These results support the effectiveness of coordinated SDG and temporal-frequency fusion under the evaluated clean and controlled synthetic degradation conditions.