USE CASE

Transport

Failure prediction on vehicles wheel bearing

Context

Our customer monitors the state of health of wheel bearing equipment. Many context-related parameters might lead to wrong prediction (false alarm). He can only produce his models after collecting data from cars during 18 months, and these data must contain failures. This is far too long expensive, and not accurate. 

Need

The customer wanted to get a model that estimated the state of health of the wheel bearing system, and a model robust to variation of context parameters (state of road, type of driving, weather conditions, …).

Solution

Using healthy data, a blind fault detector has been designed and then validated on blind new data for which labels have been removed. The resulting correlations between predicted and client labels were very satisfactory.

Results

Customer feedback: the time to produce the failure prediction model has been astonishingly short and this suggests that the deployment of our solution to many other design issues (brakes, tires, etc…) can help saving a lot of R&D engineering development time.

Illustration of the generated features’ behaviour for bad wheel bearing systems. Black and Red curves represent respectively the lower and upper bound of the features for a healthy wheel bearing system. The fact that the features go so ostensibly outside the healthy box indicates unhealthy wheel bearing system.