Work

MMDA: Masked Multi-Teacher Distillation Architecture Search for Industrial Time Series Prediction

Year: 2026
Type: article
Abstract: Industrial time series prediction is a powerful tool for prognostics and health management (PHM) of complex equipment. Although deep learning methods have shown remarkable performance, most rely on manually designed neural networks that require substantial expert knowledge and often conflict with the limited computational resources in real-world edge computing environments. To address these challenges, we propose a masked multi-teacher distillation architecture (MMDA) search method for industrial time series prediction, a lightweight neural architecture search framework based on differentiable optimization and multi-teacher self-distillation. The proposed framework incorporates a dual-objective mask strategy that jointly optimizes the architecture search by dynamically selecting informative teachers. The proposed MMDA framework is validated on both the public C-MAPSS dataset and a bearing dataset. Experimental results demonstrate that MMDA effectively discovers accurate and efficient models for industrial time series prediction, outperforming state-of-the-art algorithms, improving RMSE by 17.3% on average. Note to Practitioners—Predicting equipment failure is critical for ensuring industrial safety and reducing operational costs. However, deep learning models often require significant expertise to design and are too computational for affordable edge hardware. To address these challenges, this paper proposes an automated framework that designs predictive models which are both highly accurate and lightweight. The proposed method has been validated on two datasets, and the results indicate a significant improvement in prediction accuracy over other leading methods, making reliable health monitoring more accessible for industrial applications.
Language: English
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204134233
FWCI:
Cites: 37
Cited by:
Related to: 10