Echo State Network applications in structural health monitoring

摘要

Echo State Networks (ESNs), a type of recurrent neural network, have been applied to multi-dimensional, longitudinal, time-series data obtained from an array of sensors in the context of structural health monitoring (SHM) and structural surveys. It has been shown that ESNs are able to process both spatial and temporal data as a means of detecting structural damage in two case study applications. The first of these was for the detection of corrosion in reinforced concrete. A magnetic flux leakage (MFL) technique was employed to gather a large database of MFL signals from a reinforced concrete test bed featuring artificially inserted breaks and corrosion. An ESN was trained to recognise characteristic defect signals arising in the MFL data and was then presented with a full set of spatial test data from the test bed. A separate MFL-ESN was also trained to recognise the noise that can be seen in the end regions following the MFL energisation process. Combining the two ESNs allowed for the accurate determination of the location of defects. The second application involved data from the National Physical Laboratory's footbridge project. The bridge was embedded with ten temperature and eight tilt sensors, which took data readings at five-minute intervals over a three-year period. It was then subjected to a series of damage and repair cycles. Three separate ESN analysis approaches were used. In the first of these, an ESN (ESNa) was trained on the relationship between the temperature sensor and tilt sensor readings prior to the first damage cycle, so as to learn the bridge's normal patterns of behaviour. Presenting the trained ESN with the remaining temperature data for the full time period then allowed it to predict the eight tilt sensor readings at each remaining time step. Any significant difference between the ESN prediction of normal behaviour for each tilt sensor and the actual tilt sensor reading would therefore be indicative of an abnormal change in the state of the bridge, which might in turn be suggestive of damage. A second ESN (ESNb) was trained to detect the characteristic signals in the raw tilt sensor data at some of the exact moments when the bridge was damaged. It was found that ESNb was able to classify perfectly one type of event signal, while also proving to be highly successful at classifying both a second type of event and normal behaviour. The third ESN approach (ESNc) saw this difference used to train another ESN, whose task was to indicate permanent changes in the state of the bridge due to damage. Using this three pronged approach in this context, ESNa could be used to locate the damage on the bridge, ESNc can determine whether or not the bridge has been permanently damaged and ESNb could then pinpoint the time when the damage had occurred and the type of event causing it.