Modern condition monitoring of wind turbine shaft line relies on a wealth of fault-sensitive indicators. Due to the complexity of the mechanisms and the non-stationary operations conditions, dozens of indicators are monitored for a single machine. As a rise in the corresponding indicators is expected whenever a fault is developing in the machine, alarms are designed to be triggered whenever one indicator exceeds a threshold. Choosing an appropriate threshold is complex, with compromises made between false and missed alarms. This often complicates the work of maintenance operators where alarms need to be investigated by experienced maintenance operators to assess the gravity of the fault. Further, when it comes to the monitoring of a fleet of machines the value of the indicator itself may change from one machine to another. As such, the decision-making process does not only rely on the value of the indicator itself, but also on its dynamic evolution. Despite abundant literature on thresholding strategies, there is a lack of monitoring techniques focused on trends. There is a need to develop a robust trend analysis framework dedicated to the surveillance of rotating machinery on which the temporal behavior of the indicators is compressed into a concise form. Qualitative trend analysis [1,2] aims at decomposing a series into a succession of segments with identical qualitative properties (increasing, concave, ...etc.). The segments are labeled from a library of minimal shapes called ‘primitives’ describing the qualitative state. Feedback on run-to-failure surveillance of mechanical equipment usually describes three states of a developing fault [3]. Before the appearance of the damage, the indicator is supposed to show no trend and be stationary. Then, the incipient fault may induce a linear trend, degenerating into an exponential increase until catastrophic failure. The proposed qualitative trend analysis will stem from this schematic fault development. Segments will be described from a primitive-based language including three shapes: constant, linear, and exponential. A method is presented to segment the series into qualitative episodes with an assigned primitive. On each segment, the qualitative information can be completed with the estimated parameter of the regression. After validation on numerical examples, the method is applied to several condition indicators from an industrial damaged wind turbine.