Anomaly Detection without Structural Behavior Models

摘要

Measurement-data interpretation is especially costly when structural modeling is required. This paper summarizes results of three studies of anomaly detection without structural behavior models. In the first study, five signal-analysis methods were compared using an eight-year temperature/traffic loading vs. time history on a simulation of a two-span beam with four cases of damage scenarios occurring after four years and nine months. Moving principal-component analysis was identified to be able to detect damage most reliably. In the second study, ten signal analysis methods were compared using the same loading vs. time history as the first study. Fifteen cases of variations in daily temperature, traffic loading and sensor noise were used in this comparison. Results showed that moving principal-component analysis and robust regression detect damage more effectively than the other methods, especially in the presence of outliers and missing data. The third study involved the application of temperature cleansing methods to increase damage detectability and decrease detection time when using moving principal components analysis. The advantages of this method were compared with those of the robust regression method. Both methods are complementary. Complex relationships between frequency filtering, sensor/damage locations, traffic loading and noise were found to determine detectability and detection time. This has lead to several avenues for further study.