Fault detection in distributed parameter systems (such as thermal or fluid processes) plays a crucial role in ensuring both safety and efficiency. Current data-driven strategies often disregard the dynamic evolution of these processes, making them unsuitable for rapidly changing or transient conditions. Meanwhile, most reported model-based approaches rely heavily on the backstepping technique, which lacks sufficient redundancy for reliable fault diagnosis since only boundary measurements are utilized. To address these limitations, this work investigates the robust fault detection problem from a model-based perspective, incorporating both boundary and in-domain sensing. A real-time fault detection filter is proposed, eliminating the dependence on extensive data acquisition and offline training. Comprehensive theoretical analysis is provided to guide parameter tuning and threshold design. In particular, a time-varying threshold is constructed to suppress false alarms during transient stages. Experimental validation on a hot strip mill cooling system demonstrates the approach’s feasibility and potential for industrial deployment.