Fault Tolerant Adaptive Systems
(Faculty Mentor: W. K. Jenkins)
As
digital multimedia cable and wireless communication networks become more
stressed by heavy traffic and more highly integrated at the circuit level it is
anticipated that increasingly efficient adaptive strategies will be useful for
self-diagnosis of failure, and real-time system-level reconfiguration will
become increasingly important to maintain functional integrity. The scientific
objective of this research project is to make use of non-canonical adaptive
filter architectures to achieve improved learning characteristics for both FIR
and IIR adaptive filters that operate under conditions of permanent or transient
hardware faults.
In classical digital filter design the term canonical structure refers to an architecture that realizes a desired filter characteristic with the minimum number of delays, adders, and multipliers. Generally speaking, a canonical structure is highly desirable because it achieves the design goal with the lowest possible cost in terms of hardware components. However, it is also possible to meet design goals with non-canonical architectures, in which the system realization is over-parameterized, and hence is not a minimal cost realization. Such over-parameterization may in fact provide additional desirable properties that are not provided by canonical realizations. For example, it has been well known since the 1970’s that state space digital filters designed to have more than the minimal number of nonzero entries in the A and B matrices (non-canonical) can produce filter implementations that have lower round-off noise accumulation in the filter output than that of canonical state space structures. Adaptive Fault Tolerance (AFT) takes advantage of non-canonical architectures that rely on the inherent adaptive process as an automatic fault tolerance mechanism. Under normal operating conditions adaptive systems, such as adaptive echo cancelers, adaptive equalizers, and adaptive noise reduction filters, adjust their own system parameters to reduce a specified error criterion. Hardware failures in such systems will hamper their ability to minimize the error criterion to the greatest possible extent. However, such systems will continue to adapt their parameters to reduce this error to the greatest possible extent despite the occurrence of hardware failures. This research investigates ways to design a system that makes use of the ongoing adaptive process to automatically compensate for a wide variety of hardware failures. The main objective of the proposed research is to further explore the use of over-parameterization for achieving better performance and improved reliability in both FIR and IIR adaptive filter structures.