摘要
Modern processes rely on thousands of sensors, yet operators still lack trustworthy tools for measurement-driven fault detection and diagnosis. Deep learning excels at capturing nonlinearity and dynamics, but its opaque decision-making limits use in safety-critical processes. This survey fills that gap by introducing a measurement-oriented taxonomy of explainable neural networks (XNNs) and explainable graph neural networks (XGNNs) from the view of instrumentation and measurement. XNNs are posited as variable-centered instruments that assign calibrated importance scores to individual sensors for different faults. XGNNs are framed as topology-centric instruments, allowing direct measurement of interaction strength and causal propagation among units and control loops. This review delivers step-by-step guidelines that convert historical data into explainable detectors, trackers, and diagnostic meters. A comparison highlights when an XNN suffices and when an XGNN is mandatory, giving instrumentation engineers a decision chart. Different from prior surveys, we show that graphs are the faithful way to integrate P&IDs, material balances, and causal knowledge into deep learning measurements: XGNN explanations map directly onto process diagrams, creating on-screen instruments that display both the alarm and its physical trail. Finally, it concludes by identifying open challenges and recommending future directions for industrial deployment.