Why Hirundo?
Highlights
Access to IAV Hirundo
Contact
Inquiry

Why Hirundo?

Car manufacturers want to evaluate the condition of their car fleets in the field in order to assess the risk of individual components failing. Precise knowledge of the aging condition enables maintenance to be carried out as needed, thereby reducing the number of breakdowns.

Forecasting future failure rates offers further advantages. It allows future warranty costs to be estimated and spare parts to be better planned. IAV’s patented probabilistic life cycle prediction method addresses precisely these issues.

The method enables the prediction of future failures by analyzing vehicle-specific mileage and current failure statistics. At the heart of the process is the use of probabilistic models that combine the advantages of machine learning and statistics. This allows all uncertainties in the data to be taken into account and made visible for evaluation.

Complete automation eliminates time-consuming steps of data cleansing and model adjustment, making it easy to use even for non-experts. The method is implemented in Microsoft Azure as a machine learning pipeline and has been validated in many use cases. It is currently being implemented as a software-as-a-service application in AWS.

Automatic Data Cleansing

The imported fleet and failure data is cleaned up in several steps and prepared so that it can be processed reliably. Not only invalid and duplicate values are removed, but also non-monotonic entries. The process has been robustified using many real-world test cases, so that this step can be performed completely automatically. This saves users the time-consuming task of filtering individual data manually.

Automatic Model Selection

An optimal model structure is automatically sought for the available failure data in order to determine the actual number of failed components. To this end, various model candidates are trained using the expectation maximization algorithm and compared with each other. The optimal structure is determined on the basis of statistical criteria, so that no manual intervention is necessary. This makes the application possible even for non-experts.

Vehicle-Specific Mileage Estimate

Probabilistic life expectancy forecasting determines an individual aging model for each vehicle. This enables a much more accurate estimate of the mileage distribution of a fleet than is possible with the widely used three-point estimate. As probabilistic methods, the underlying aging models are able to represent future aging as a probability distribution. Depending on the data available, either vehicle-specific data or fleet information is used for this purpose.

Consideration of Uncertainties

The method is based on the consistent use of probabilistic models, both for estimating mileage and for calculating the probability of failure. The uncertainties underlying the data are taken into account within both models and offset against each other. As a result, the method not only provides a single point estimate, but also the overall probability of occurrence.

Prediction of Failures

When looking at the service life curve alone, it is only possible to make a general statement about the number of failures at a certain mileage. In reality, however, vehicles are driven differently, and there is no fleet in which all vehicles have the same mileage. To predict the number of failures, probabilistic life expectancy forecasting therefore considers the mileage of individual vehicles and projects this into the future in order to infer future failures from the failure risks of individual vehicles.

Experience IAV Hirundo:

IAV Hirundo on the AWS Marketplace
Login to IAV Hirundo
Contact

Dr. Jan Gieseler

jan.gieseler@​iav.de LinkedIn

Dominik Gütermann

dominik.guetermann@​iav.de LinkedIn

Contact