PSI - Issue 57

Jeroen Van Wittenberghe et al. / Procedia Structural Integrity 57 (2024) 95–103 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

101

7

Fig. 8: Comparison between numerical longitudinal strains extracted along the FBG sensor lines and experimental strain data from the two FBG sensor lines. The two peaks observed in the numerical data are associated with the position of the loads (trolley wheels). Du e to the hook position, the load is not evenly distributed over the two trolley wheel pairs. Therefore, the peaks have different heights. 3.3. Fatigue analysis A series of simulations were performed to calibrate a neural network model based on Radial Basis Function (RBF). After the neural network is calibrated, stresses at any particular crane model point can be obtained given as input the hook load and trolley position (Fig. 9). The stresses along the welds are further used in fatigue analysis. Such an approach is beneficial as no finite element simulations are further required after the neural network calibration. A similar methodology was applied by Zuo et al. (2023) to perform fatigue assessments in a tower crane. The resulting approach for the current SHM system is illustrated in Fig. 10. The SCADA data and FBG data on the physical twin (the actualcrane)is used in several steps of the digital twin analysis. Firstly, the measurement data is used for the digital twin model calibration. Once the digital twin is calibrated, the neural netwo rk approach is used to derive functions for each weld that calculate stress. Similarly a function is generated that predicts the strain at the sensor locations using the SCADA data as input. The first set of functions is used for the further fatigue analysis of the steel structure based on EN 13001. By performing rainflow counting, the remaining life can be calculated for all the welds of the crane. The latter function that calculates the expected strain sensor values based on the SCADA data can be used for detection of events that differ from the expected strain response. E.g. through such analysis the occurrence of horizontal bending can be identified and quantified automatically. The system data is further processed and visualised in a user interface that visualises the results from the remaining life assessment together with statistical data on the usage of the crane. Part of this output is illustrated in Fig. 11 where the load histogram with the distribution of the loads is illustrated together with a visual representation of all evaluated welds in the crane. Also the areas with maximum fatigue damage are illustrated.

Fig. 9: Neural network schematic where represents weights between the RBF neurons and the output node.