REPORT BRIEF
Fuzzy Learning In Engineering Diagnosis
April 9, 1998
CENTER FOR ENGINEERING EDUCATION AND PRACTICE
SCHOOL OF ENGINEERING, UNIVERSITY OF MICHIGAN-DEARBORN
 
AUTHOR(S): Yi Lu Murphey 
Department of Electrical and Computer Engineering
PARTNER(S): Brennan Hamilton1, Lee Feldkamp2 
1Advanced Vehicle Technology - Core and Product Engineering, Ford Motor Company, Dearborn, MI 48121. 
2Ford Research Laboratory, Ford Motor Company, Dearborn, MI 48121.
ASSISTANT(S): Tie-Qi Chen 
Research assistant
 
 
BACKGROUND Fault diagnosis has been a classic engineering problem. Particularly, automotive fault diagnosis has the following difficulties: 
  • Knowledge on most vehicle diagnostic problems is incomplete and vague due to the complexity of the modern vehicles.
  • Different vehicle models have different engineering features, e.g. Ford ThunderBird is very different from Ford Town Car. Furthermore, ThunderBird 96 may be different from ThunderBird 97.
  • Often the data available for learning is not complete, it is important for the fuzzy system to combine human expert knowledge with the knowledge learned from data samples.
  • When the vehicle model changes, for example, vehicles of a new model are manufactured, we have very few data available to build the fuzzy knowledge base. It is important for the system to have the capability of accumulating learning.
  • Data samples of good vehicles and vehicles having problems are often unbalanced. The data acquired at an assembly plant generally contains 99% good vehicles and 1% vehicles that have the problem to be diagnosed.
OBJECTIVES
     
  • To solve the problems outlined above utilizing fuzzy system techniques.
  • To develop an intelligent fuzzy diagnostic system for the End-of-Line vehicle test in assembly plants.
  • To create an intelligent fuzzy system that can automatically generate the required knowledge base including the fuzzy rules.
APPROACH We applied fuzzy logic and machine learning techniques to solve this problem. 
  • We have developed a fuzzy intelligent system for EEC vacuum leak diagnosis.
  • The system has 2 major components:
    1. learning engineering diagnostic knowledge
      • automatic fuzzy rule generation
      • automatic fuzzy membership function optimization
    2. engineering fault diagnosis
  • Learning diagnosis knowledge
 
  • Fuzzy Rule Generation
    • Fuzzy rules generated from engineering heuristic knowledge
      • engineers generate fuzzy rules through an interactive linguistic interface
      • engineers modify the fuzzy rules generated by the system
    • Automatic fuzzy rule generation from training data
 
  • Membership function optimization
    • MSF’s are represented by triangular functions
    • MSF optimization is to find the best locations and overlapping of the triangular functions
    • Optimization methods investigated:
      • gradient descent method
      • stochastic annealing method
  • Fuzzy engineering diagnosis
 
RESULTS An algorithm has been developed for automatic fuzzy rule generation. 
  • An intelligent fuzzy system has been developed and implemented for diagnosing EEC vacuum leak. The system is capable of learning the characteristics of different vehicle models and performing robust test. The system is in the process of being implemented into Ford Motor Company's diagnostic systems.
  • The fuzzy diagnostic system was tested on data sets of over 20,000 samples, the accuracy of the system was above 96%. We tested the system on two different types of vehicle models:
    • Ford Thunderbird 95 and 96 models
    • Ford Lincoln Towncar 96, 97 and 98 models
  • As a result of this project, we received a grant from the National Science Foundation to support a project titled "A Distributed Fuzzy System Model for Automotive Diagnosis." The duration of the project is two years and the award is in the amount of $93,450.00. We have also received funding from the Ford Motor Company in the amount of $40,000.00 for a period of two years.
CONCLUSIONS We have developed and implemented a fuzzy model for automotive fault diagnosis. The fuzzy model has two modes, learning and detection. Within the learning mode, a new rule generation method and a membership function optimization method are used. The model has been implemented into a fuzzy diagnostic system that detects the vacuum leak in Electronic Engine Controller in automobiles. The fuzzy vacuum leak diagnostic system has been tested on two different vehicle models, Thunderbird and Lincoln Towncar. The testing results show that the system is effective, fast and compact, suitable for running on a PC platform.
IMPACT
     
  • Educational impact:
    • A course based on this project, "Fuzzy Diagnosis in Automotive Engineering," was offered to Ford engineers in May 1996. There were 20 participants in this course.
    • The result is also used in senior students project design, graduate courses including ECE 579 Intelligent Systems and ECE 580 Neural Networks.
  • Industrial impact: The result of this project is currently being transferred to a Ford diagnostic program to be used in Ford assembly plants.
 This page is prepared by Jie Chen, May 11, 1998.