Vibration-Based Predictive Maintenance in Manufacturing: A Systematic Review of Signal Processing, Artificial Intelligence, and Industrial Deployment

Abstract

Vibration-based Predictive Maintenance (PdM) will prove to be one of the most significant developments in industrial asset management in the last 20 years. This systematic review provides an overview of mechanical vibration analysis evolution from simple periodic inspection and calendar-based scheduling to the current deep learning-supported edge deployment of diagnostic architectures with Industry 4.0 goals. The review summarizes the signal processing techniques, such as the Short-Time Fourier Transform (STFT), Empirical Mode Decomposition (EMD), wavelet packet transforms, cyclostationary analysis, and their combinations, and illustrates their linkage with machine learning classifiers and deep learning networks, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, transformer networks, Graph Neural Networks (GNNs), and Physics-Informed Neural Networks (PINNs). Significant implementation issues in the central challenges are discussed in detail: data scarcity, class imbalance, non-stationarity, edge resource constraints, model interpretability, and cybersecurity concerns in IIoT deployments. While several emerging mitigation strategies such as transfer learning, generative adversarial augmentation, TinyML quantisation, transparent operator networks, and federated learning are explored, there are still gaps that can be evaluated. Lastly, there are six priority research directions are identified: zero-shot diagnostics, digital twin enhanced training, foundation model integration, adversarially resilient pipelines, federated multi-plant learning, and sustainable sensor lifecycle management. The review shows how making full use of autonomous prescriptive maintenance requires computational efficiency, physical interpretability, and uniform ethical oversight.