PGP-FD: A Lightweight Phase-Guided Generative Graph Process for Resource-Efficient Edge Fault Pattern Recognition in Fire-Protection Valve Sensing Systems

Edge fire-protection sensing systems must recognize latent valve faults under limited computation, memory, and energy budgets, although fault samples are scarce and operational responses are strongly phase-dependent. This paper proposes PGP-FD, a lightweight phase-guided generative graph process for event-driven edge fault pattern recognition. Each valve operation or inspection is represented as a compact response window. A depthwise temporal encoder captures local dynamics, a monotonic phase gate models the ordered evolution of standby, actuation, transient, recovery, and settling phases, and a compact measurement graph learns hydraulic-mechanical coupling among pressure, valve motion, and event context. To reduce class imbalance, a fault-conditioned stochastic process generates rare events during training under range, phase-order, and valve-pressure consistency constraints. The training-only posterior and offline synthesis branches are removed during deployment, while confidence-uncertainty prefix inference avoids unnecessary continuous computation and retains full processing for difficult cases. Experiments on real fire-protection valve data show that PGP-FD improves recognition quality, response speed, and online model compactness over representative baselines. The results demonstrate a practical integration of lightweight pattern recognition and mechanism-consistent generative learning for resource-efficient industrial edge intelligence.