The dataset was made using four vibrations sensors placed in different directions.

The load applied to the gearbox varied between 0 and 90 percent of its maximum.

We then reformated the data to have 4 x 256 elements per lines.

We were able to distinguish between the 6 following conditons:

• Healthy gearbox at 10 percent load
• Healthy gearbox at 40 percent load
• Healthy gearbox at 90 percent load
• Gearbox with a broken tooth at 10 percent load
• Gearbox with a broken tooth at 40 percent load
• Gearbox with a broken tooth at 90 percent load

With 95% accuracy, this proves to be a simple and cheap solution to a very common and costly issue in predictive maintenance.