Automatic detection of process faults requires process expertise to determine the fault symptoms and mathematical algorithms that can classify these symptoms correctly. Thus, the effectiveness of any automatic detection system can degrade because of inconsistencies in the process data and variations in the assumed symptoms of the fault. This paper presents a novel detection strategy based on the data characterization and reliability analysis. In more detail, at first process data are profiled to automatically select and apply the most suitable detection algorithms. Then, the reliability of the detection decisions made by the selected detection algorithms is evaluated, considering a decision unreliable when the process data fit neither the healthy case nor the faulty case assumed by the method. The strategy is applied to a valve stiction detection system. In addition, an exponential fitting method, recognizing three oscillation patterns associated with stiction, is proposed and incorporated into the system. The resulting system was tested on a benchmark data, and the results are discussed.