Department of Mathematics

Ford Corporation

Title: Product defect diagnosis
.
Ford Research Contact:

Dr. Michael Cavaretta
Technology Leader
Infotronics and System Analytics
Ford Research and Advanced Engineering
Ford Motor Company
voice: (313) 594 - 4733
email: mcavaret@ford.com
Mike leads the Data Mining research group at Ford Research and Advanced Engineering.

Abstract:

Background
Ford Motor Company uses hazard analysis to identify part / component / assembly failure rates over time from warranty claims data. This project challenges you to diagnose the causes behind those failure rates through decision tree analysis. The new analysis is intended to reveal where / when / how failures happen from the defective hardware warranty claims.

Technical Description
To support this analysis at Ford Motor Company, you will develop software that implements hazard analysis within a decision tree framework. The software will use Weibul, or another well known failure distribution, to project the failure rate of part/component/assembly failure rates over time for an unlimited number of attributes. The attributes will be generated from Ford's Analytical Warranty System.

The software should be capable of implementing the following requirements:

1. Input a comma separated file containing vehicle records. Each record represents one vehicle and each field represents one attribute associated with that vehicle. The software should be able to handle an arbitrary number of records and fields.
2. Rank ordering of the important vehicle attributes with regard to the failure in question. That is, where it is attempting to diagnose failures of starter motors, it should calculate the importance of each variable in determining subpopulations with the highest failure rate. This calculation should be done using hazard analysis, not the entropy measure used in standard decision tree software. The calculation should be fast enough so that 100 attributes can be calculated in less than five seconds.
3. The rank ordering of important variables should replace the standard entropy value in building a complete decision tree. The results of a decision tree should be graphical, and preferably interactive by the user.
4. Java and a Web browser interface are the preferred methods of implementation for the software. If necessary, other code may be used for performance reasons.

Business Results
This software will provide a new ability for failure analysis at Ford Motor Company. If successful, an implementation in one of the quality systems would provide users with the capability to quickly analyze large quantities of warranty data.