Automated Learning Applied to Fault Diagnosis

Automated learning methods can be used to design fault diagnosis procedures. When the characteristics of the measurements that distinguish the various faults are unknown, they can be ``learned´´ from example measurements on faulty systems. A learning algorithm is presented for determining which of several possible faults exists in a system. The procedure is demonstrated on a system where the test conditions preclude the use of traditional diagnosis procedures. When applied to actual hardware, the experimental results show good agreement with the theoretical limit of diagnosability. The resulting diagnosis is faster, simpler, and requires fewer measurements than other methods.