Autocovariance Functions as Damage Sensitive Features: Case Study of a Lattice Tower

Many tower structures, such as wind turbines, are continuously monitored to detect damage at an early stage. Due to their symmetrical cross-section, tower structures often exhibit closely spaced modes, making system identification challenging and often subject to significant uncertainties. Other damage-sensitive features (DSFs), such as autocovariance functions (ACFs), are currently being researched as an alternative to modal parameters. ACFs can be quickly and easily estimated from the acceleration measurement data without the need for system identification. Since ACFs have the same mathematical form as a free decay of the system under stationary random excitation, they carry information about the natural frequencies and mode shapes. In this paper, ACFs are used as DSFs in combination with a whitening transformation for data normalisation. This method does not require measurements of environmental conditions (ECs). A disadvantage of the investigated method is the tendency to result in false negatives when the structure is weakly excited. To address this issue, an extension of the method that reduces the excitation dependence of the DSFs is proposed in this contribution. While the method’s ability to detect and localise damage has been demonstrated using simulated and experimental data, it has not yet been systematically tested for different damage scenarios. Therefore, in this paper, the method is applied to an ambient-excited lattice tower featuring reversible damage mechanisms. Six damage scenarios are investigated here, differing in location and severity. With the extended method, damage can be reliably detected in five of the six investigated damage scenarios, showing the applicability of ACFs for damage detection. However, challenges persist, notably false positives arising from unrepresented ECs in the training data. This finding underscores the importance of encompassing a wide range of ECs in the training phase to mitigate such false alarms. These results underline that physical interpretability and data normalisation play an important role in selecting suitable DSFs. Damage localisation using the proposed method is compared to the established mode shape curvature method. Both methods are only able to correctly localise three of the investigated damages. However, the results differ significantly when the damage cannot be localised correctly.