Self-adaptive software for signal processing

Publisher Website
Google Scholar
Abstract
Digital signal processing (DSP) systems are widely used in communication, medical, sonar, radar, equipment health monitoring and many other applications. Frequently, the signal processing system has to meet real-time requirements and provide very large throughput. For example, modern automatic target recognition systems operate with a processing throughput in excess of 10 Gflop per second. In real-time vibration analysis used for turbine engine testing (1), the aggregate sustained computation rate is also in the Gflop range. The high performance requires the use of computing platforms that include the combination of dedicated hardware processors, and general-purpose computers forming a hybrid, parallel/distributed configuration. Complexity, heterogeneity of the computing environment, and real-time operation make the software development for digital signal processing difficult and expensive. The design of DSP systems is based on the available a priori information about the signal source and the noise in the environment. Necessarily, the performance of the implemented system largely depends on the environmental conditions as well as additional factors, such as the computational resources available to the task. In conventional design approaches, these conditions are typically set in design time by introducing various constraints, simplifications, and assumptions. The critical issue in this methodology is what happens if the design time assumptions do not hold? Stabilization of the environment is impossible in many applications. For example, in turbine engine testing, the deterioration of sensors is unavoidable. The goal of designing a single, sufficiently robust DSP algorithm, which is able to tolerate any possible changes in the environment, is also not practical. The price of robustness is usually decreased performance or increased complexity, which is undesirable. The most attractive approach is to design a system, which is able to adapt to the changing circumstances. Adaptive DSP algorithms are almost exclusively designed for processing systems with fixed, predefined structure, where adaptation occurs solely by means of parameter adjustments. By its very nature, the fixed- structure approach has strong limitations in abruptly changing, unstable environments. While parameter adaptive systems (PAS) have existed and have been widely applied, the use of structural adaptation is a newly emerging research field (2). A structurally adaptive system (SAS) is able to modify its own structure (i.e. the composition of the signal flow) while it is running. Structurally adaptive digital signal processing raises complex problems from the point of view of computation as well as system dynamics. This paper addresses the computational aspect of structural adaptation. We show an approach that demonstrates how an adaptive software architecture can be created and used to solve signal-processing problems.

This publication has 4 references indexed in Scilit: