Multi-Scale Decoupling of Industrial Dynamics Via Trend-Fluctuation Interaction-Aware Transformer for Quality Prediction

Abstract

Accurate prediction of key quality variables is crucial for monitoring and optimizing modern industrial processes. However, most existing methods remain constrained by single-scale modeling, making it difficult to capture long-term global trends and short-term local fluctuations simultaneously. In addition, the dynamic couplings between these multi-scale components are often overlooked, leading to insufficient feature extraction. To address these limitations, a multi-scale trend-fluctuation interaction-aware transformer (MTI-Former) is proposed in this paper. First, a decoupling layer based on discrete wavelet transform (DWT) is designed to decompose industrial data into low-frequency trend and high-frequency fluctuation signals. Then, an adaptive high-pass enhancement filter is introduced to amplify critical high-frequency details and improve the perception of local disturbances. Cross-scale coupling is modeled through a trend-fluctuation interaction-aware attention module, which captures dynamic interactions between trends and fluctuations. Subsequently, a trend-fluctuation decoupling attention module applies a dual-path cross-attention mechanism to separately extract global dependencies and local variations. Finally, a gating mechanism fuses these representations to generate comprehensive multi-scale temporal predictions. The effectiveness of MTI-Former is verified on two real industrial datasets, and extensive results show that it outperforms several state-of-the-art methods in industrial quality prediction.