An interpretable hybrid adaptive framework combining empirical mode decomposition and deep learning for real-time nonstationary signal analysis in smart manufacturing

Purpose In the context of Industry 4.0, smart manufacturing systems produce complex, nonstationary and nonlinear signal data streams that challenge traditional analytical methods. This study aims to develop an interpretable hybrid adaptive framework for accurate, real-time analysis of such signals, addressing limitations in adaptability, interpretability and deployment in mission-critical environments. Design/methodology/approach The framework integrates empirical mode decomposition (EMD) to decompose raw multivariate sensor data into intrinsic mode functions (IMFs) without a fixed basis, capturing localized temporal behaviors. These IMFs are then processed via a hybrid deep learning architecture combining convolutional neural networks (CNNs) for spatial feature extraction, long short-term memory (LSTM) units for temporal pattern learning, and an attention mechanism for interpretability and decision transparency. Validation involved experiments on benchmark datasets from real-world smart manufacturing settings, including multi-sensor vibration and thermal signals from CNC machines. Findings The proposed approach outperformed state-of-the-art models in predictive accuracy and adaptability to signal drift, achieving over 10% improvement in classification F1-score and significantly reduced detection latency. Attention-based visualization enhanced model interpretability, facilitating practical deployment on edge devices. Originality/value This work addresses a critical gap in adaptive signal analysis for nonstationary processes by contributing a novel, scalable, explainable and edge-compatible artificial intelligence framework that bridges signal decomposition with deep learning, suitable for next-generation industrial systems.