MS18.6 - System Identification and Damage Detection

. Vibration analysis is highly beneficial in a variety of engineering areas. However, in many real-world applications, vibration data acquisition may be challenging due to the accessibility of the desired sensor locations. It can be also costly if many measurement points are required. Consequently, a few vibration estimation methods have been proposed, which are referred to as “virtual sensing”. Virtual sensing claims to be able to replace a physical sensor with a virtual one, whose signal should closely resemble the signal from the physical sensor if it was placed at the same location. The signal from such a virtual sensor is estimated based on a numerical model of the structure under test and data from a number of physical sensors. In this study, the sensitivity of virtual sensing performance to physical sensors location is explored. The well-known modal expansion and decomposition-based virtual sensing method is examined. Two scenarios are considered: (i) the most common scenario where the physical sensors are placed in the nodes of a regular mesh, and (ii) where the physical sensors configuration is generated by the optimal sensors placement (OSP) algorithm. The experimental examination is performed on a simple test structure (rectangular aluminum plate) using time and frequency domain performance indicators for three excitation profiles (pseudo-random, burst pseudo-random, and sinusoidal). The results demonstrate that the use of OSP significantly improves the methodology’s performance.

The comfort and stability of maglev train depends on the suspension gap between bogie and rail, which should be strictly restricted in a certain range, usually 8 to 10 millimeters, by using the suspension controller. However, this restriction may become unstable due to the influence of external disturbance like the suspension controller failure. When such failure appears, the electromagnetic force will suddenly change, lead to the suspension gap out of balance, and even cause a collision between bogie and rail. To keep the operation of maglev train, this study aims to analyze and identify the maglev suspension controller failure. With a group of accelerometers installed on rail, the acceleration data that represents the structural dynamic response of rail, can be collected to observe the failure pattern of maglev suspension controller. In general, the failure pattern is analyzed by the dynamic characteristics of acceleration data from both perspective of time and frequency domain. Then, to realize the intelligent identification of such failure, a method based on convolution neural network is developed to distinguish the normal and failed suspension controller. It is found that there is a significant high frequency component at a certain moment when the suspension controller fails, while this phenomenon is absent in a normal suspension controller. Based on this discriminative feature, the convolution neural network enables the accurate binary classification on suspension controller failure. As a result, the findings from this study are expected to condition monitoring on suspension controller.

Time-dependent phenomena, such as creep of concrete, induce stress/strain redistributions in post-tensioned concrete bridges that must be taken into account during their design. The underestimation of time-dependent effects can negatively influence the load-deformation behavior of this bridge typology, leading to unexpected deformations of the deck, serviceability issues, and collapses. Numerical models able to simulate time-dependent phenomena should also be used to properly conduct the structural assessment of existing post-tensioned concrete bridges. However, before the introduction of specific materials’ constitutive laws to account for time-dependent phenomena, numerical models should be calibrated and validated according to the information gathered by testing the structures they represent, to carry out reliable numerical simulations. The paper presents the preliminary results obtained from the structural assessment of a post-tensioned concrete box girder bridge with vertically prestressed internal structural joints, which exhibits marked deflections of the box girder. A numerical model of the case study bridge was constructed, hence non-destructive and partially destructive tests were carried out on the structure to characterize its structural response. Overall, the resulting numerical model represents a solid basis for conducting more advanced analyses, such as those investigating the influence of time-dependent phenomena on the long-term structural response of the structure.

This work present an innovative technique that helps to accelerate structural dynamics simulations when there is a defect in the structure and at the same time, allows its approximation in a parametric way by using reduced models. These developments open the door to the introduction of robust hybrid twins applied to structural health monitoring (SHM), where a physics-based numerical model is enriched with real measured data to correctly predict a real phenomenon, which in this case would be the dynamic response of a damaged structure. This hybrid twin is then used to train a neural network, which is able to predict the position of the damage and its severity on the structure. Several numerical examples are provided in this work, which illustrate the performance of the proposed methodology for the parameterization of dynamic simulations on a plate as well as for the identification of the location and severity of damage on it.