QED - Quantified Evolution of Degradation in gears (QED)

Condition monitoring of gears is extremely important in a number of industrial applications including wind turbines and automotive. A plethora of diagnostics indicators, based on signal processing and machine learning, applied on indirect measurements (e.g. acceleration, sound, rotational speed, etc.) have been proposed to monitor the damage evolution in gears. The thresholds for damage detection are nowadays mainly based on statistics or on ISO norms and are often too conservative. In some specific high value applications (e.g. wind turbines, aircraft & industrial gas turbines, helicopters) wear debris sensors are used to detect and quantify gear faults but the effectiveness of these sensors in early fault detection is under question, moreover they have a high cost and a high installation complexity. Vision inspections (e.g. using a video endoscope) are sometimes used for periodic inspection in machineries such as wind turbines for which the operation needs to be stopped and secured. Moreover, often, only parts of the gears are visually accessible by the inspection ports. Automatic visual inspection techniques have been developed at lab scale, by taking photos prior and after test runs ,  or by using a high-speed camera in a dedicated set-up. Their application in industrial settings requires translation to more complex gear systems. Furthermore, in the current state of arts / state of practice, the monitoring algorithms, based on visual or indirect measurements, are tuned to a specific gear system and often assume specific working regimes. Scaling these algorithms to other working regimes and / or other gear systems is not yet well established.

Due to the above limitations, the effective link between the initialization of degradation (which can be visually observed) and the detection and evolution based on indirect measurements in, and their scaling to, complex gear systems is not yet clear. Therefore, the innovation goals of the QED project are:

• Capture accurately defects initiation and degradation
• Improve the accuracy of indirect measurements
• Scaling models through transfer learning