Abstract: Aiming at the operational scenarios of "dual carbon" requirements and unit flexibility retrofitting, this study investigates the development of a data-driven predictive maintenance strategy for power and thermal control instruments, forming a closed-loop from data acquisition and preprocessing to fault warning modeling, decision-making, and resource orchestration. On the architectural side, cross-system data is unified to a 1-second timestamp, with quality bit governance and operating condition mapping implemented. On the modeling side, random forests are employed to learn features related to temperature deviation, pressure fluctuations, vibration spectrum energy, and oxygen perturbations, outputting warning probabilities and health indices. On the decision-making side, a multi-dimensional weighted approach incorporating carbon emission penalties, thermal consumption increments, and downtime costs generates maintenance packages and scheduling plans. Using a coastal power plant in East China as a case study, the results demonstrate that this strategy effectively stabilizes measurement links, reduces overall costs and emissions, and provides reusable processes and definition standards for scalable deployment