AI-Driven Condition Monitoring and Predictive Maintenance of Fluid Power Systems: A Review of Diagnostic Models and Applications

Abstract

Hydraulic and pneumatic machinery, collectively referred to as fluid power systems, underpin a wide range of industrial automation, heavy-equipment, and aerospace actuation applications. Yet these systems remain susceptible to gradual degradation processes, including seal wear, cavitation, fluid contamination, and valve erosion, any of which can trigger unplanned downtime and inflate maintenance expenditure if not caught early. Classical model-based diagnostic techniques generally struggle to represent the nonlinear, stochastic character of fault progression in such machinery. This paper reviews a decade of research on artificial-intelligence-driven condition monitoring and predictive maintenance developed specifically for fluid power systems. It examines both supervised and unsupervised learning strategies, spanning convolutional neural networks (CNNs), long short-term memory (LSTM) networks, autoencoders, support vector machines (SVMs), and more recent transformer-based architectures, as applied to fault detection, fault classification, and remaining useful life (RUL) estimation. Published diagnostic models are assessed against a consistent set of benchmarks, including accuracy, false positive rate, real-time deployability, and sensor requirements, to enable a fair, cross-study comparison. The resulting analysis shows that hybrid deep-learning architectures consistently surpass single-method approaches, reaching fault-classification accuracies above 93% under controlled experimental conditions. The review further highlights persistent gaps in the field, most notably the limited ability of existing models to adapt across differing operating conditions and the continuing shortage of industry-validated, publicly available datasets. The paper closes with a discussion of promising future directions, including physics-informed neural networks and edge-deployable AI architectures.