Robust damage detection and localization algorithms are critical to the development of practical structural health monitoring systems. Damage decisions are typically based on time series measurements of structural responses. Determining the correct signal processing and classification algorithms for a particular problem represents a difficult task requiring detailed knowledge of the structure, sensor system, and damage cases of interest. Typically, a low-dimensional feature vector is extracted from the time series measurements and used as input to a classification or regression system for decision making. The difficulty lies in selecting the signal processing and feature extraction steps to generate an optimal feature vector for a specific application.

Evolutionary computational methods such as genetic algorithms and swarm intelligence have been used to solve a wide range of applied problems across many engineering disciplines. One method in particular, genetic programming, has previously shown promise for automated feature extraction from speech and image signals. Genetic programming evolves a population of candidate solutions represented as computer programs to optimize a user-defined fitness function. Importantly, genetic programming allows both the size and structure of the solution to evolve without any upfront decisions by the user. Current work is exploring how genetic programming can be applied to design of feature extraction processes, excitation and actuation signals, and other components of SHM systems.

With the Autofead method, a custom genetic programming variant performs automated feature extraction algorithm design for a training database of numeric sequence inputs. Possible input representations include time series, spectral measurements, frequency response functions, and mode shape estimates. Preliminary results indicate significant performance benefits over traditional SHM damage features on a variety of laboratory-scale benchmarking structures. Further development of the Autofead method will evaluate and improve on the robustness and generalization of evolved solutions.