Different from traditional health-monitoring methods based on vibrational signals recorded by contact sensors, an online diagnosis procedure for steel truss structures using sound signals was proposed. The basic idea of the procedure was to identify the features related to bolt connection damage extracted from sound signals and locate the damaged position. Before the online diagnosis was carried out, sound signals were specifically collected by a microphone array involving environmental noise and sound discharged by artificial damaged bolt connections. Then the signals were preprocessed and their time and frequency domain features were extracted, from which sensitive features were selected by support vector machine recursive feature elimination. A support vector machine classifier aiming to identify signals related to damage was trained with the selected sensitive features, and a genetic algorithm was used to optimize its parameters. An improved method called steered response power and phase transformation with offline database was put forward to compute the steered response power values of coordinates in the offline database to localize the source of identified damage signals. The pre-built database consisted of a series of coordinates indicating the positions of bolts. When the online diagnosis was implemented for a steel truss structure, sound signals were picked up by the microphone array at the same location as that used for the database construction. The signals were preprocessed and their sensitive features were extracted for damage identification by the trained support vector machine classifier. If some signals were judged to be related to bolt connection damage, steered response power and phase transformation with offline database was used to compute steered response power values, with which a fusion decision was made based on evidence theory to locate the damaged bolt connection. The experiment of a steel truss model with 24 bolt connections showed that the proposed procedure could locate the loose bolts precisely even under heavy noise effect, and had a smaller computational load compared with traditional steered response power and phase transformation.
ANOMALY DETECTION FROM CAHNGES IN STATISTICAL PATTERNS OF STRUCTURAL DYNAMIC CHARACTERISTICS OF A SIMPLY SUPPORTED STEEL-TRUSS BRIDGE