Time-Series Forecasting in Industrial Environments: A Performance Study and a Novel Late Fusion Framework

In manufacturing environments, monitoring of the overall equipment effectiveness (OEE) via soft sensors plays a pivotal role in enhancing productivity and efficiently planning maintenance schedules. However, the accurate forecasting of the OEE presents considerable challenges due to the complexity of manufacturing data and equipment interdependence across stages. To this end, advanced time-series forecasting methods based on deep learning (DL) pose a promising avenue in tackling these challenges. In this study, we present a taxonomy of DL forecasting architectures, consisting of multilayer perceptrons (MLPs), recurrent models, Transformer-based models, and temporal convolutional networks (TCNs), and we perform a comparative study of the state-of-the-art approaches. Additionally, a lightweight late fusion linear architecture is proposed, incorporating patching, moving average (MA) decomposition, and Fourier Transform decomposition (PDFLinear), and an exponentially weighted averaging (EWA) module responsible for late fusion. Representative state-of-the-art models of each taxonomy class are benchmarked using a real-world antenna assembly line use case and compared against our proposed method. The experimental results show that our proposed model consistently matches or outperforms the state-of-the-art models in terms of forecasting efficacy for all forecast horizons, while requiring a fraction of the computational resources.